BigML.io—The BigML API
Documentation
Overview
Last Updated: Thursday, 2018-10-18 13:19
This page provides an introduction to BigML.io—The BigML API. A quick start guide for the impatient is here.
BigML.io is a Machine Learning REST API to easily build, run, and bring predictive models to your project. You can use BigML.io for basic supervised and unsupervised machine learning tasks and also to create sophisticated machine learning pipelines.
BigML.io is a REST-style API for creating and managing BigML resources programmatically. That is to say, using BigML.io you can create, retrieve, update and delete BigML resources using standard HTTP methods.
BigML.io gives you:
- Secure programmatic access to all your BigML resources.
- Fully white-box access to your datasets, models, clusters and anomaly detectors.
- Asynchronous creation of resources.
- Near real-time predictions.
Jump to:
- BigML Resources
- REST API
- HTTPS
- Base URL
- Version
- Summary of Resource URL Patterns
- Summary of HTTP Methods
- Resource ID
- Libraries
- Limits
BigML Resources
BigML.io gives you access to the following resources: project, source, dataset, sample, correlation, statisticaltest, configuration, and composite.
The four original BigML resources are: source, dataset, model, and prediction.
As shown in the picture below, the most basic flow consists of using some local (or remote) training data to create a source, then using the source to create a dataset, later using the dataset to create a model, and, finally, using the model and new input data to create a prediction.
The training data is usually in tabular format. Each row in the data represents an instance (or example) and each column a field (or attribute). These fields are also known as predictors or covariates.
When the machine learning task to learn from training data is supervised one of the columns (usually the last column) represents a special attribute known as objective field (or target) that assigns a label (or class) to each instance. The training data in this format is named labeled and the machine learning task to learn from is named supervised learning.
Once a source is created, it can be used to create multiple datasets. Likewise, a dataset can be used to create multiple models and a model can be used to create multiple predictions.
A model can be either a classification or a regression model depending on whether the objective field is respectively categorical or numeric.
Often an ensemble (or collection of models) can perform better than just a single model. Thus, a dataset can also be used to create an ensemble instead of a single model.
A dataset can also be used to create a cluster or an anomaly detector. Clusters and Anomaly Detectors are both built using unsupervised learning and therefore an objective field is not needed. In these cases, the training data is named unlabeled.
A centroid is to a cluster what a prediction is to a model. Likewise, an anomaly score is to an anomaly detector what a prediction is to a model.
There are scenarios where generating predictions for a relative big collection of input data is very convenient. For these scenarios, BigML.io offers batch resources such as: batchprediction, batchcentroid, and batchanomalyscore. These resources take a dataset and respectively a model (or ensemble), a cluster, or an anomaly detector to create a new dataset that contains a new column with the corresponding prediction, centroid or anomaly score computed for each instance in the dataset.
When dealing with multiple projects, it's better to keep the resources that belong to each project separated. Thus, BigML also has a resource named project that helps you group together all the other resources. As you will see, you just need to assign a source to a pre-existing project and all the subsequent resources will be created in that project.Note: In the snippets below you should substitute Alfred's username and API key for your own username and API Key.
REST API
BigML.io conforms to the design principles of Representational State Transfer (REST). BigML.io is entirely HTTPS-based.
You can create, read, update, and delete resources using the respective standard HTTP methods: POST, GET, PUT and DELETE.
All communication with BigML.io is JSON formatted except for source creation. Source creation is handled with a HTTP PUT using the "multipart/form-data" content-type.
HTTPS
All access to BigML.io must be performed over HTTPS. In this way communication between your application and BigML.io is encrypted and the integrity of traffic between both is verified.
Base URL
All BigML.io HTTP commands use the following base URL:
https://bigml.ioVersion
The BigML.io API is versioned using code names instead of version numbers. The current version name is "andromeda" so URLs for this version can be written to require this version as follows: https://bigml.io/andromeda/Specifying the version name is optional. If you omit the version name in your API requests, you will always get access to the latest API version. While we will do our best to make future API versions backward compatible it is possible that a future API release could cause your application to fail.
Specifying the API version in your HTTP calls will ensure that your application continues to function for the life cycle of the API release.
Summary of Resource URL Patterns
BigML.io gives you access to the following resources: project, source, dataset, sample, correlation, statistical test, configuration, and composite.
https://bigml.io/project
https://bigml.io/source
https://bigml.io/dataset
https://bigml.io/sample
https://bigml.io/correlation
https://bigml.io/statisticaltest
https://bigml.io/configuration
https://bigml.io/compositeSummary of HTTP Methods
BigML.io uses the standard POST, GET, PUT, and DELETE HTTP methods to create, retrieve, update and delete resources, respectively.| Operation | HTTP method | Semantics |
|---|---|---|
| CREATE | POST | Creates a new resource. Only certain fields are "postable". This method is not idempotent. Each valid POST request results in a new directly accessible resource. |
| RETRIEVE | GET | Retrieves either a specific resource or a list of resources. This methods is idempotent. The content type of the resources is always "application/json; charset=utf-8". |
| UPDATE | PUT | Updates partial content of a resource. Only certain fields are "putable". This method is idempotent. |
| DELETE | DELETE | Deletes a resource. This method is idempotent. |
Resource ID
All BigML resources are identified by a name composed of two parts separated by a slash "/". The first part is the type of the resource and the second part is a 24-char unique identifier. See the examples below:
source/4f510d2003ce895676000069
dataset/4f510cfc03ce895676000040
model/4f51473203ce89b7ef000005
ensemble/523e9017035d0772e600b285
prediction/4f51473b03ce89b7ef000008
evaluation/50a30a453c19200bd1000839Libraries
We have developed light-weight API bindings for Python, Node.js, and Java.
A number of libraries for many other languages have been developed by the growing BigML community: C#, Ruby, PHP , and iOS. If you are interested in library support for a particular language let us know. Or if you are motivated to develop a library, we will give you all the support that we can.
Limits
BigML.io is currently limited to 1,000,000 (one million) requests per API key per hour. Please email us if you have a specific use case that requires a higher rate limit.Sources
Last Updated: Thursday, 2018-10-18 13:19
A source is the raw data that you want to use to create a predictive model. A source is usually a (big) file in a comma separated values (CSV) format. See the example below. Each row represents an instance (or example). Each column in the file represents a feature or field. The last column usually represents the class or objective field. The file might have a first row named header with a name for each field.
Plan, Talk,Text,Purchases,Data,Age,Churn?
family, 148, 72, 0, 33.6, 50, TRUE
business, 85, 66, 0, 26.6, 31, FALSE
business, 83, 64, 0, 23.3, 32,TRUE
individual, 9, 66, 94, 28.1, 21, FALSE
family, 15, 0, 0, 35.3, 29, FALSE
individual, 66, 72, 175, 25.8, 51,TRUE
business, 0, 0, 0, 30, 32, TRUE
family, 18, 84, 230, 45.8, 31,TRUE
individual, 71, 110, 240, 45.4, 54, TRUE
family, 59, 64, 0, 27.4, 40, FALSEA source:
- Should be a comma-separated values (CSV) file. Spaces, tabs, semicolons and tabs are also valid separators.
- Weka's ARFF files are also supported.
- JSON in a few formats is also supported. See below for more details.
- Microsoft Excel files or Mac OS numbers files should also work most times. But it would be better if you please export them to CSV (commad-separated values).
- Cannot be bigger than 64GB.
- Can be gzipped (.gz) or compressed (.bz2). It can be zipped (.zip), but only if the archive contains one single file.
You can also create sources from remote locations using a variety of protocols like https, hdfs, s3, asv, odata/odatas, dropbox, gcs, gdrive, etc. See below for more details.
BigML.io allows you to create, retrieve, update, delete your source. You can also list all of your sources.
Jump to:
- JSON Sources
- Source Base URL
- Creating a Source
- Creating a Source Using a Local File
- Creating a Source Using a URL
- Creating a Source Using Inline Data
- Creating a Source with Automatically Generated Synthetic Data
- Text Processing
- Items Detection
- Datetime Detection
- Source Arguments
- Retrieving a Source
- Source Properties
- Filtering and Paginating Fields from a Source
- Updating a Source
- Deleting a Source
- Listing Sources
JSON Sources
BigML.io can parse JSON data in one of two formats:
- A top-level list of lists of atomic values, each one defining a row:
Valid JSON Source format[ ["sepal length","sepal width","petal length","petal width","species"], [5.1,3.5,1.4,0.2,"Iris-setosa"], [4.9,3.0,1.4,0.2,"Iris-setosa"], ... ] - A top-level list of dictionaries, where each dictionary's values represent the row values and the corresponding keys the column names. The first dictionary defines the keys that will be selected.
Valid JSON Source format[ {"sepal length":5.1,"sepal width":3.5,"petal length":1.4,"petal width":0.2 ,"species":"Iris-setosa"}, {"sepal length":4.9,"sepal width":3.0,"petal length":1.4,"petal width":0.2 ,"species":"Iris-setosa"}, ... ]
Source Base URL
You can use the following base URL to create, retrieve, update, and delete sources. https://bigml.io/sourceAll requests to manage your sources must use HTTPS and be authenticated using your username and API key to verify your identity. See this section for more details.
Creating a Source
You can create a new source in any of the following four ways:
- Local Sources: Using a local file. You need to post the file content in "multipart/form-data". The maximum size allowed is 64 GB per file.
- Remote Sources: Using a URL that points to your data. The maximum size allowed is 64 GB or 5 TB if you use a file stored in Amazon S3.
- Inline Sources: Using some inline data. The content type must be "application/json". The maximum size in this case is limited 10 MB per post.
- Synthetic Sources: Automatically generate synthetic data sources, presumably for activities such as testing, prototyping, and benchmarking.
Creating a Source Using a Local File
To create a new source, you need to POST the file containing your data to the source base URL. The file must be attached in the post as a file upload.The Content-Type in your HTTP request must be "multipart/form-data" according to RFC2388. This allows you to upload binary files in compressed format (.Z, .gz, etc) that will be uploaded faster.
You can easily do this using curl. The option -F (--form) lets curl emulate a filled-in form in which a user has pressed the submit button. You need to prefix the file path name with "@".
curl "https://bigml.io/source?$BIGML_AUTH" -F file=@iris.csvCreating a Source Using a URL
To create a new remote source you need a URL that points to the data file that you want BigML to download for you.
You can easily do this using curl. The option -H lets curl set the content type header while the option -X sets the http method. You can send the URL within a JSON object as follows:
curl "https://bigml.io/source?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"remote": "https://static.bigml.com/csv/iris.csv"}'You can use the following types of URLs to create remote sources:
- HTTP or HTTPS. They can also include basic realm authorization.
Example URLshttps://test:test@static.bigml.com/csv/iris.csv http://archive.ics.uci.edu/ml/machine-learning-databases/iris/iris.data - Public or private files in Amazon S3.
Example Amazon S3 URLss3://bigml-public/csv/iris.csv s3://bigml-test/csv/iris.csv?access-key=AKIAIF6IUYDYUQ7BALJQ&secret-key=XgrQV/hHBVymD75AhFOzveX4qz7DYrO6q8WsM6ny
Creating a remote source from Google Drive and Google Storage
You have two options to create a remote datasource from Google Drive and Google Storage via API:
- Using BigML:
Allow BigML to access to your Google Drive or Google Storage from the Cloud Storages section from your Account or from your Dashboard sources list. You will get the access token and the refresh token.
Google Drive example:- Select the option to create source from Google Drive:
- Allow BigML access to your Google Drive:
- Get the access token and refresh token:
You can easily create the remote source using curl as in the examples below:You should complete steps above in the same way to import source from Google Cloud Storage.
> Creating a remote source from Google Drivecurl "https://bigml.io/source?$BIGML_AUTH" \ -X POST \ -H 'content-type: application/json' \ -d '{"remote":"gdrive://noserver/0BxGbAMhJezOScTFBUVFPMy1xT1E?token=ya29.AQGn2MO578Fso0fVF0hGb0Q60ILCagAwvFEZoiovK5kvPWSt5z2QMbXDyAvqUtQeYdJbA39YkyRq3A&refresh-token=1/00qQ1040yDSyh_0DRLYZRnkC_62gE8M5Tpb68sfvmj8"}'
> Creating a remote source from Google Cloud Storagecurl "https://bigml.io/source?$BIGML_AUTH" \ -X POST \ -H 'content-type: application/json' \ -d '{"remote":"gcs://company_bucket/Iris.csv?token=ya29.AQGn2MO578Fso0fVF0hGb0Q60ILCagAwvFEZoiovK5kvPWSt5z2QMbXDyAvqUtQeYdJbA39YkyRq3A&refresh-token=1/00qQ1040yDSyh_0DRLYZRnkC_62gE8M5Tpb68sfvmj8"}'
- Using your own app:
You can also create a remote source from your own App. You first need to authorize BigML access from your own Google Apps application. BigML only needs authorization for read-only authentication scope (
https://www.googleapis.com/auth/devstorage.read_only,https://www.googleapis.com/auth/drive.readonly), but you can have any of the other available scopes (find authentication scopes available for Google Drive and Google Storage). After the authorization process you will get your access token and refresh token from the Google Authorization Server.
Then the process is the same as creating a remote source using BigML application described above. You need to POST to the source endpoint an object containing at least the file ID (for Google Drive) or the bucket and the file name (for Google Storage) and the access token, but in this case you will also need to include the app secret and app client from your App. Again, including the refresh token is optional.
Your values for app client and app secret appear as Client secret and Client ID in Google developers console respectively. See image below.
> Creating a remote source from Google Drive using your appcurl "https://bigml.io/source?$BIGML_AUTH" \ -X POST \ -H 'content-type: application/json' \ -d '{"remote":"gdrive://noserver/0BxGbAMhJezOSXy1oRU5MSU90SUU?token=ya29.AQGn2MO578Fso0fVF0hGb0Q60ILCagAwvFEZoiovK5kvPWSt5z2QMbXDyAvqUtQeYdJbA39YkyRq3A&refresh-token=1/00qQ1040yDSyh_0DRLYZRnkC_62gE8M5Tpb68sfvmj8&app-secret=AvFake1Secretjt27HQWTm4h&app-client=667300000007-07gjg5o912o1v422hfake2cli3nt3no6.apps.googleusercontent.com"}'
> Creating a remote source from Google Cloud Storage using your appcurl "https://bigml.io/source?$BIGML_AUTH" \ -X POST \ -H 'content-type: application/json' \ -d '{"remote":"gcs://company_bucket/Iris.csv?token=ya29.AQGn2MO578Fso0fVF0hGb0Q60ILCagAwvFEZoiovK5kvPWSt5z2QMbXDyAvqUtQeYdJbA39YkyRq3A&refresh-token=1/00qQ1040yDSyh_0DRLYZRnkC_62gE8M5Tpb68sfvmj8&app-secret=AvFake1Secretjt27HQWTm4h&app-client=667300000007-07gjg5o912o1v422hfake2cli3nt3no6.apps.googleusercontent.com"}'
Creating a Source Using Inline Data
You can also create sources sending some inline data within the body of a POST http request. This way is specially useful if you want to model small amounts of data generated by an application.
To create an inline source using curl you can use the following example:
curl "https://bigml.io/source?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"data": "a,b,c,d\n1,2,3,4\n5,6,7,8"}'Independently of how you create a new source (local, remote or inline) BigML.io will return a newly created source document, if the request succeeded.
{
"category": 0,
"code": 201,
"content_type": "application/octet-stream",
"created": "2012-11-15T02:24:59.686739",
"credits": 0.0,
"description": "",
"disable_datetime": false,
"fields_meta": {
"count": 0,
"limit": 200,
"offset": 0,
"total": 0
},
"file_name": "iris.csv",
"md5": "d1175c032e1042bec7f974c91e4a65ae",
"name": "iris.csv",
"number_of_datasets": 0,
"number_of_models": 0,
"number_of_predictions": 0,
"private": true,
"project": null,
"resource": "source/4f603fe203ce89bb2d000000",
"size": 4608,
"source_parser": {},
"status": {
"code": 1,
"message": "The request has been queued and will be processed soon"
},
"tags": [],
"type": 0,
"updated": "2012-11-15T02:24:59.686758"
}
Creating a Source with Automatically Generated Synthetic Data
You can also synthetically create sources using automatically generated data for activities such as testing, prototyping, and benchmarking.
To create a syntheric source using curl you can use the following example:
curl "https://bigml.io/source?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"synthetic": {"fields": 10, "rows": 10}}'In addition to the file, you can also use the following arguments.
Source Arguments
| Argument | Type | Description |
|---|---|---|
|
category
optional |
Integer, default is 0 |
The category that best describes the data. See the category codes for the complete list of categories.
Example: 1 |
|
data
optional |
String |
Data for inline source creation.
Example: "a,b,c,d\n1,2,3,4\n5,6,7,8" |
|
description
optional |
String |
A description of the source up to 8192 characters long.
Example: "This is a description of my new source" |
|
disable_datetime
optional |
Boolean, default is false |
Whether BigML has to generate or not new fields from existing date-time fields.
Example: true |
|
file
optional |
multipart/form-data; charset=utf-8 | File containing your data in csv format. It can be compressed, gzipped, or zipped if the archive contains only one file |
|
item_analysis
optional |
Object, default is shown in the table below |
Set of parameters to activate item analysis for the source.
Example: |
|
name
optional |
String, default is Unnamed source |
The name you want to give to the new source.
Example: "my new source" |
|
project
optional |
String |
The project/id you want the source to belong to.
Example: "project/54d98718f0a5ea0b16000000" |
|
remote
optional |
String |
A URL pointing to file containing your data in csv format. It can be compressed, gzipped, or zipped.
Example: https://static.bigml.com/csv/iris.csv |
|
source_parser
optional |
Object, default is shown in the table below |
Set of parameters to parse the source.
Example: |
|
synthetic
optional |
Object, default is shown in the table below |
Set of parameters to generate a synthetic source presumably for activities such as testing, prototyping and benchmarking.
Example: |
|
tags
optional |
Array of Strings |
A list of strings that help classify and index your source.
Example: ["best customers", "2018"] |
|
term_analysis
optional |
Object, default is shown in the table below |
Set of parameters to activate text analysis for the source.
Example: |
|
webhook
optional |
Object |
A webhook url and an optional secret phrase. See the Section on Webhooks for more details.
Example: |
A source parser object is composed of any combination of the following properties.
You can also use curl to customize your new source with a name and different parser. For example, to create a new source named "my source", without a header and with "x" as the only missing token.
curl "https://bigml.io/source?$BIGML_AUTH" \
-F file=@iris.csv \
-F 'name=my source' \
-F 'source_parser={"header": false, "missing_tokens":["x"]}'If you do not specify a name, BigML.io will assign to the source the same name as the file that you uploaded. If you do not specify a source_parser, BigML.io will do its best to automatically select the parsing parameters for you. However, if you do specify it, BigML.io will not try to second-guess you.
A item_analysis object is composed of any combination of the following properties.
A term_analysis object is composed of any combination of the following properties.
| Property | Type | Description |
|---|---|---|
|
bigrams
optional |
Boolean, default is false |
Whether to include a contiguous sequence of two items from a given sequence of text. See n-gram for more information.This argument is deprecated in favor of ngrams and is equivalent to ngrams=2.
Example: true DEPRECATED |
|
case_sensitive
optional |
Boolean, default is false |
Whether text analysis should be case sensitive or not.
Example: true |
|
enabled
optional |
Boolean, default is true |
Whether text processing should be enabled or not.
Example: true |
|
excluded_terms
optional |
Array of Strings, default is [], an empty list. |
Specifies a list of terms to ignore when performing term analysis.
Example: |
|
language
optional |
String, default is "en" |
The default language of text fields in a two-letter language code, which will change the resulting stemming and tokenization. Available options are: "ar", "ca", "cs", "da", "de", "en", "es", "fa", "fi", "fr", "hu", "it", "ja", "ko", "nl", "pl", "pt", "ro", "ru", "sv", "tr", "zh", "none", or null for auto-detect.
Example: "es" |
|
ngrams
optional |
Integer, default is 1 |
A positive integer n that specifies the use of all sequences of consecutive tokens of length n should be considered as terms, in addition to their constituent tokens (when separated by a single space and no stopwords). See n-gram for more information.The minimum value is 1 and maximum value is 5.
Example: 5 |
|
stem_words
optional |
Boolean, default is true |
Whether lemmatization (stemming) of terms should be done, according to linguistic rules in the provided language. Note that if the language is, for example, zh, even English words will not be lemmatized as with English rules.
Example: true |
|
stopword_diligence
optional |
String, default is "light" |
The aggressiveness of stopword removal, where the levels are light, normal or aggressive, in order, where each level is a superset of words in the previous ones. The most common languages will add stopwords at each level, but less common languages may not.
Example: "normal" |
|
stopword_removal
optional |
String, default is "selected_language" |
A string or keyword specifying the type of stopword removal to perform. Available options are where it can be none (remove no stopwords), selected_language (remove stopwords from the provided language), and all_languages (remove stopwords from all languages). Note that this parameter supersedes use_stopwords if provided. Also note that the null language does have a non-empty stopword list such as single numeric digits.
Example: "all_languages" |
|
term_filters
optional |
Array of Strings |
Filters that should be applied to the chosen terms. Available options are:
Example: "html_keywords" |
|
term_regexps
optional |
Array | A list of strings specifying regular expressions to be matched against input documents. If present, these regular expressions will automatically be chosen for the final term list, and their per-document occurrence counts will be the number of matches of the expression in that document. |
|
token_mode
optional |
String, default is "all" |
Whether tokens_only, full_terms_only or all should be tokenized.
Example: "tokens_only" |
|
use_stopwords
optional |
Boolean, default is true |
Whether to use stop words or not. This fields is deprecated in favor of stopword_removal.
Example: true DEPRECATED |
A synthetic object is composed of the following properties.
Text Processing
While the handling of numeric, categorical, or items fields within a decision tree framework is fairly straightforward, the handling of text fields can be done in a number of different ways. BigML.io takes a basic and reasonably robust approach, leverging some basic NLP techniques along with a simple bag-of-words style method of feature generation.
At the source level, BigML.io attempts to do basic language detection. Initially the language can be English ("en"), Spanish ("es"), Catalan/Valencian ("ca"), Dutch ("nl"), French ("fr"), German ("de"), Portuguese ("pt"), or "none" if no language is detected. In the near future, BigML.io will support many more languages.
For text fields, BigML.io adds potentially five keys to the detected fields, all of which are placed in a map under term_analysis.
The first is language, which is mapped to the detected language.
There are also three boolean keys, case_sensitive, use_stopwords, and stem_words. The case_sensitive key is false by default. use_stopwords should be true if we should include stopwords in the vocabulary for the detected field during text summarization. stem_words should be true if BigML.io should perform word stemming on this field, which maps forms of the same term to the same key when summarizing or generating models. By default, use_stopwords is false and stem_words is true for languages other than "none" and they are not present otherwise.
Finally, token_mode determines the tokenization strategy. It may be set as either tokens_only, full_terms_only, and all. When set as tokens_only then individual words are used as terms. For example, "ML for all" becomes ["ML", "for", "all"]. However, when full_terms_only is selected, then the entire field is treated as a single term as long as it is shorter than 256 characters. In this case "ML for all" stays ["ML for all"]. If all is selected, then both full terms and tokenized terms are used. In this case ["ML for all"] becomes ["ML", "for", "all", "ML for all"]. The default for token_mode is all.
There are a few details to note:
- If full_terms_only is selected, then no stemming will occur even if stem_words is true.
- Also, when either all or tokens_only are selected, a term must appear at least twice to be selected for the tag cloud. However full_terms_only lowers this limit to a single occurrence.
- Finally, if the language is "none", or if a language does not have an algorithm available for stopword removal or stemming, the use_stopwords and stem_words keys will have no effect.
Items Detection
BigML automatically detects as items fields that have many different categorical values per instance separated by non-alphanumeric characters, so they can’t be considered either categorical or text fields
These kind of fields can be found in transactional datasets where each instance is associated to a different set of products contained within one field. For example, datasets containing all products bought by users or prescription datasets where each patient is associated to different treatments. These datasets are commonly used for Association Discovery to find relationships between different items.
Find the two CSV examples below that could be considered items fields:
User, Prescription
John Doe, medicine 1; medicine 2
Jane Roe, medicine 1; medicine 3; medicine 4; medicine 6
Transaction, Product
12345, product 1; product 2; product 5; product 6; product 7
67890, product 1; product 3; product 4
In the examples above, the fields Prescription and Products will be considered as items and each different value will be a unique item.
Once a field has been detected as items, BigML tries to automatically detect which is the best separator for your items. For example, for the following itemset {hot dog; milk, skimmed; chocolate}, the best separator is the semicolon which yields three different items: 'hot dog', 'milk, skimmed' and 'chocolate'.
For items fields, there are five different parameters you can configure under the property group item_analysis, which includes separator that allows you to specify which separator you want to set for your items.
Note that items fields can’t be eligible as target fields for models, logistic regression, and ensembles, but they can be used as predictors. For anomaly detection, they can’t be included as an input field to calculate the anomaly score, although they can be selected as summary fields.
Datetime Detection
During the source pre-scan BigML tries to determine the data type of each field in your file. This process automatically detects datetime fields and, if disable_datetime is not explicitly set to "false", BigML will generate additional fields with their components.
For instance, if a field named "date" has been identified as a datetime with format "YYYY-MM-dd", four new fields will be automatically added to the source, namely "date.year", "date.month", "date.day-of-month" and "date.day-of-week". For each row, these new fields will be filled in automatically by parsing the value of their parent field, "date". For example, if the latter contains the value "1969-07-14", the autogenerated columns in that row will have the values 1969, 7, 14 and 1 (because that day was Monday). As noted before, autogenaration can be disabled by setting disable_datetime option to "true", either in the create source request or later in an update source operation.
When a field is detected as datetime, BigML tries to determine its format for parsing the values and generate the fields with their components. By default, BigML accepts ISO 8601 time formats (YYYY-MM-DD) as well as a number of other common European and US formats, as seen in the table below:
| time_format Name | Example |
|---|---|
| basic-date-time | 19690714T173639.592Z |
| basic-date-time-no-ms | 19690714T173639Z |
| basic-ordinal-date-time | 1969195T173639.592Z |
| basic-ordinal-date-time-no-ms | 1969195T173639Z |
| basic-t-time | T173639.592Z |
| basic-t-time-no-ms | T173639Z |
| basic-time | 173639.592Z |
| basic-time-no-ms | 173639Z |
| basic-week-date | 1969W297 |
| basic-week-date-time | 1969W297T173639.592Z |
| basic-week-date-time-no-ms | 1969W297T173639Z |
| clock-minute | 5:36 PM |
| clock-minute-nospace | 5:36PM |
| clock-second | 5:36:39 PM |
| clock-second-nospace | 5:36:39PM |
| date | 1969-07-14 |
| date-hour | 1969-07-14T17 |
| date-hour-minute | 1969-07-14T17:36 |
| date-hour-minute-second | 1969-07-14T17:36:39 |
| date-hour-minute-second-fraction | 1969-07-14T17:36:39.592 |
| date-hour-minute-second-ms | 1969-07-14T17:36:39.592 |
| date-time | 1969-07-14T17:36:39.592Z |
| date-time-no-ms | 1969-07-14T17:36:39Z |
| eu-date | 14/7/1969 |
| eu-date-clock-minute | 14/7/1969 5:36 PM |
| eu-date-clock-minute-nospace | 14/7/1969 5:36PM |
| eu-date-clock-second | 14/7/1969 5:36:39 PM |
| eu-date-clock-second-nospace | 14/7/1969 5:36:39PM |
| eu-date-millisecond | 14/7/1969 17:36:39.592 |
| eu-date-minute | 14/7/1969 17:36 |
| eu-date-second | 14/7/1969 17:36:39 |
| eu-ddate | 14.7.1969 |
| eu-ddate-clock-minute | 14.7.1969 5:36 PM |
| eu-ddate-clock-minute-nospace | 14.7.1969 5:36PM |
| eu-ddate-clock-second | 14.7.1969 5:36:39 PM |
| eu-ddate-clock-second-nospace | 14.7.1969 5:36:39PM |
| eu-ddate-millisecond | 14.7.1969 17:36:39.592 |
| eu-ddate-minute | 14.7.1969 17:36 |
| eu-ddate-second | 14.7.1969 17:36:39 |
| eu-sdate | 14-7-1969 |
| eu-sdate-clock-minute | 14-7-1969 5:36 PM |
| eu-sdate-clock-minute-nospace | 14-7-1969 5:36PM |
| eu-sdate-clock-second | 14-7-1969 5:36:39 PM |
| eu-sdate-clock-second-nospace | 14-7-1969 5:36:39PM |
| eu-sdate-millisecond | 14-7-1969 17:36:39.592 |
| eu-sdate-minute | 14-7-1969 17:36 |
| eu-sdate-second | 14-7-1969 17:36:39 |
| hour-minute | 17:36 |
| hour-minute-second | 17:36:39 |
| hour-minute-second-fraction | 17:36:39.592 |
| hour-minute-second-ms | 17:36:39.592 |
| mysql | 1969-07-14 17:36:39 |
| no-t-date-hour-minute | 1969-7-14 17:36 |
| odata-format | /Datetime(-14752170831)/ |
| ordinal-date-time | 1969-195T17:36:39.592Z |
| ordinal-date-time-no-ms | 1969-195T17:36:39Z |
| rfc822 | Mon, 14 Jul 1969 17:36:39 +0000 |
| t-time | T17:36:39.592Z |
| t-time-no-ms | T17:36:39Z |
| time | 17:36:39.592Z |
| time-no-ms | 17:36:39Z |
| timestamp | -14718201 |
| timestamp-msecs | -14718201000 |
| twitter-time | Mon Jul 14 17:36:39 +0000 1969 |
| twitter-time-alt | 1969-7-14 17:36:39 +0000 |
| twitter-time-alt-2 | 1969-7-14 17:36 +0000 |
| twitter-time-alt-3 | Mon Jul 14 17:36 +0000 1969 |
| us-date | 7/14/1969 |
| us-date-clock-minute | 7/14/1969 5:36 PM |
| us-date-clock-minute-nospace | 7/14/1969 5:36PM |
| us-date-clock-second | 7/14/1969 5:36:39 PM |
| us-date-clock-second-nospace | 7/14/1969 5:36:39PM |
| us-date-millisecond | 7/14/1969 17:36:39.592 |
| us-date-minute | 7/14/1969 17:36 |
| us-date-second | 7/14/1969 17:36:39 |
| us-sdate | 7-14-1969 |
| us-sdate-clock-minute | 7-14-1969 5:36 PM |
| us-sdate-clock-minute-nospace | 7-14-1969 5:36PM |
| us-sdate-clock-second | 7-14-1969 5:36:39 PM |
| us-sdate-clock-second-nospace | 7-14-1969 5:36:39PM |
| us-sdate-millisecond | 7-14-1969 17:36:39.592 |
| us-sdate-minute | 7-14-1969 17:36 |
| us-sdate-second | 7-14-1969 17:36:39 |
| week-date | 1969-W29-7 |
| week-date-time | 1969-W29-7T17:36:39.592Z |
| week-date-time-no-ms | 1969-W29-7T17:36:39Z |
| weekyear-week | 1969-W29 |
| weekyear-week-day | 1969-W29-7 |
| year-month | 1969-07 |
| year-month-day | 1969-07-14 |
It might happen that BigML is not able to determine the right format of your datetime field. In that case, it will be considered either a text or a categorical field. You can override that assignment by setting the optype of the field to datetime and passing the appropriate format in time_formats. For instance:
curl "https://bigml.io/source/4f603fe203ce89bb2d000000?$BIGML_AUTH" \
-X PUT \
-d '{"fields": {"000004": {"optype": "datetime", "time_formats": ["date"]}}}' \
-H 'content-type: application/json'
curl "https://bigml.io/source/4f603fe203ce89bb2d000000?$BIGML_AUTH" \
-X PUT \
-d '{"fields": {"000004": {"optype": "datetime", "time_formats": ["YYYY-MM-dd"]}}}' \
-H 'content-type: application/json'Retrieving a Source
Each source has a unique identifier in the form "source/id" where id is a string of 24 alpha-numeric characters that you can use to retrieve the source.
To retrieve a source with curl:
curl "https://bigml.io/source/4f603fe203ce89bb2d000000?$BIGML_AUTH"You can also use your browser to visualize the source using the full BigML.io URL or pasting the source/id into the BigML.com dashboard.
Source Properties
Once a source has been successfully created it will have the following properties.
| Property | Type | Description |
|---|---|---|
|
category
filterable, sortable, updatable |
Integer | One of the categories in the table of categories that help classify this resource according to the domain of application. |
| code | Integer | HTTP status code. This will be 201 upon successful creation of the source and 200 afterwards. Make sure that you check the code that comes with the status attribute to make sure that the source creation has been completed without errors. |
|
content_type
filterable, sortable |
String | This is the MIME content-type as provided by your HTTP client. The content-type can help BigML.io to better parse your file. For example, if you use curl, you can alter it using the type option "-F file=@iris.csv;type=text/csv". |
|
created
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the source was created with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
|
credits
filterable, sortable |
Float | The number of credits it cost you to create this source. |
|
description
updatable |
String | A text describing the source. It can contain restricted markdown to decorate the text. |
|
disable_datetime
updatable |
Boolean | False when BigML didn't generate new fields from existing date-time fields. |
|
fields
updatable |
Object |
A dictionary with an entry per field (column) in your data. Each entry includes the column number, the name of the field, the type of the field, a specific locale if it differs from the source's one, and specific missing tokens if the differ from the source's one. This property is very handy to update sources according to your own parsing preferences.
Example: |
| fields_meta | Object | A dictionary with meta information about the fields dictionary. It specifies the total number of fields, the current offset, and limit, and the number of fields (count) returned. |
|
file_name
filterable, sortable |
String | The name of the file as you submitted it. |
| md5 | String | The file MD5 Message-Digest Algorithm as specified by RFC 1321. |
|
name
filterable, sortable, updatable |
String | The name of the source as your provided or the name of the file by default. |
|
number_of_anomalies
filterable, sortable |
Integer | The current number of anomalies that use this source. |
|
number_of_anomalyscores
filterable, sortable |
Integer | The current number of anomaly scores that use this source. |
|
number_of_associations
filterable, sortable |
Integer | The current number of associations that use this source. |
|
number_of_associationsets
filterable, sortable |
Integer | The current number of association sets that use this source. |
|
number_of_centroids
filterable, sortable |
Integer | The current number of centroids that use this source. |
|
number_of_clusters
filterable, sortable |
Integer | The current number of clusters that use this source. |
|
number_of_correlations
filterable, sortable |
Integer | The current number of correlations that use this source. |
|
number_of_datasets
filterable, sortable |
Integer | The current number of datasets that use this source. |
|
number_of_ensembles
filterable, sortable |
Integer | The current number of ensembles that use this source. |
|
number_of_forecasts
filterable, sortable |
Integer | The current number of forecasts that use this source. |
|
number_of_logisticregressions
filterable, sortable |
Integer | The current number of logistic regressions that use this source. |
|
number_of_models
filterable, sortable |
Integer | The current number of models that use this source. |
|
number_of_optimls
filterable, sortable |
Integer | The current number of OptiMLs that use this source. |
|
number_of_predictions
filterable, sortable |
Integer | The current number of predictions that use this source. |
|
number_of_statisticaltests
filterable, sortable |
Integer | The current number of statistical tests that use this source. |
|
number_of_timeseries
filterable, sortable |
Integer | The current number of time series that use this source. |
|
number_of_topicdistributions
filterable, sortable |
Integer | The current number of topic distributions that use this source. |
|
number_of_topicmodels
filterable, sortable |
Integer | The current number of topic models that use this source. |
|
private
filterable, sortable |
Boolean | Whether the source is public or not. |
|
project
filterable, sortable, updatable |
String | The project/id the resource belongs to. |
| remote | String | URL of the remote data source. |
| resource | String | The source/id. |
|
shared
filterable, sortable |
Boolean | Whether the source is shared using a private link or not. |
| shared_hash | String | The hash that gives access to this source if it has been shared using a private link. |
| sharing_key | String | The alternative key that gives read access to this source. |
|
size
filterable, sortable |
Integer | The number of bytes of the source. |
|
source_parser
updatable |
Object | Set of parameters to parse the source. |
| status | Object | A description of the status of the source. It includes a code, a message, and some extra information. See the table below. |
|
subscription
filterable, sortable |
Boolean | Whether the source was created using a subscription plan or not. |
| synthetic | Object | Set of parameters to generate a synthetic source presumably for activities such as testing, prototyping and benchmarking. |
|
tags
filterable, updatable |
Array of Strings | A list of user tags that can help classify and index this resource. |
|
term_analysis
updatable |
Object | Set of parameters that define how text analysis should work for text fields. |
|
type
filterable, sortable |
Integer |
The type of source.
|
|
updated
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the source was updated with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
| webhook | Object | A webhook url and an optional secret phrase. See the Section on Webhooks for more details. |
Source Fields
The property fields is a dictionary keyed by an auto-generated id per each field in the source. Each field has as a value an object with the following properties:
For fields classified with optype "text", the default values specified in the term_analysis at the top-level of the source are used.
Non-provided flags by term_analysis take their default value, i.e., false for booleans, none for language.
Besides these global default values, which apply to all text fields (and potential text fields, such as categorical ones that might overflow to text during dataset creation), it's possible to specify term_analysis flags on a per-field basis.
For fields classified with optype "items", the default values specified in the item_analysis at the top-level of the source are used.
Like term_analysis, non-provided flags by item_analysis take their default value and it's possible to specify item_analysis flags on a per-field basis as well at the global level, too.
Source Status
Before a source is successfully created, BigML.io makes sure that it has been uploaded in an understandable format, that the data that it contains is parseable, and that the types for each column in the data can be inferred successfully. The source goes through a number of states until all these analyses are completed. Through the status field in the source you can determine when the source has been fully processed and is ready to be used to create a dataset. These are the fields that a source's status has:
| Property | Type | Description |
|---|---|---|
| code | Integer | A status code that reflects the status of the source creation. It can be any of those that are explained here. |
| elapsed | Integer | Number of milliseconds that BigML.io took to process the source. |
| message | String | A human readable message explaining the status. |
Once a source has been successfully created, it will look like:
{
"category": 0,
"code": 200,
"content_type": "application/octet-stream",
"created": "2012-11-15T02:24:59.686000",
"credits": 0.0,
"description": "",
"fields": {
"000000": {
"column_number": 0,
"name": "sepal length",
"optype": "numeric",
"order": 0
},
"000001": {
"column_number": 1,
"name": "sepal width",
"optype": "numeric",
"order": 1
},
"000002": {
"column_number": 2,
"name": "petal length",
"optype": "numeric",
"order": 2
},
"000003": {
"column_number": 3,
"name": "petal width",
"optype": "numeric",
"order": 3
},
"000004": {
"column_number": 4,
"name": "species",
"optype": "categorical",
"order": 4
}
},
"fields_meta": {
"count": 5,
"limit": 200,
"offset": 0,
"total": 5
},
"file_name": "iris.csv",
"md5": "d1175c032e1042bec7f974c91e4a65ae",
"name": "iris.csv",
"number_of_datasets": 0,
"number_of_models": 0,
"number_of_predictions": 0,
"private": true,
"project": null,
"resource": "source/4f603fe203ce89bb2d000000",
"size": 4608,
"source_parser": {
"header": true,
"locale": "en_US",
"missing_tokens": [
"",
"N/A",
"n/a",
"NULL",
"null",
"-",
"#DIV/0",
"#REF!",
"#NAME?",
"NIL",
"nil",
"NA",
"na",
"#VALUE!",
"#NULL!",
"NaN",
"#N/A",
"#NUM!",
"?"
],
"quote": "\"",
"separator": ","
},
"status": {
"code": 5,
"elapsed": 244,
"message": "The source has been created"
},
"tags": [],
"type": 0,
"updated": "2012-11-15T02:25:00.001000"
}
Filtering and Paginating Fields from a Source
A source might be composed of hundreds or even thousands of fields. Thus when retrieving a source, it's possible to specify that only a subset of fields be retrieved, by using any combination of the following parameters in the query string (unrecognized parameters are ignored):
Since fields is a map and therefore not ordered, the returned fields contain an additional key, order, whose integer (increasing) value gives you their ordering. In all other respects, the source is the same as the one you would get without any filtering parameter above.
The fields_meta field can help you paginate fields. Its structure is as follows:Updating a Source
To update a source, you need to PUT an object containing the fields that you want to update to the source' s base URL. The content-type must always be: "application/json". If the request succeeds, BigML.io will return with an HTTP 202 response with the updated source.
For example, to update a source with a new name and a new locale you can use curl like this:
curl "https://bigml.io/source/4f603fe203ce89bb2d000000?$BIGML_AUTH" \
-X PUT \
-d '{"name": "a new name", "source_parser": {"locale": "es-ES"}}' \
-H 'content-type: application/json'
Deleting a Source
To delete a source, you need to issue a HTTP DELETE request to the source/id to be deleted.
Using curl you can do something like this to delete a source:
curl -X DELETE "https://bigml.io/source/4f603fe203ce89bb2d000000?$BIGML_AUTH"If the request succeeds you will not see anything on the command line unless you executed the command in verbose mode. Successful DELETEs will return "204 no content" responses with no body.
Once you delete a source, it is permanently deleted. That is, a delete request cannot be undone. If you try to delete a source a second time, or a source that does not exist, you will receive a "404 not found" response.
However, if you try to delete a source that is being used at the moment, then BigML.io will not accept the request and will respond with a "400 bad request" response.
See this section for more details.Listing Sources
To list all the sources, you can use the source base URL. By default, only the 20 most recent sources will be returned. You can see below how to change this number using the limit parameter.
You can get your list of sources directly in your browser using your own username and API key with the following links.
https://bigml.io/source?$BIGML_AUTHDatasets
Last Updated: Thursday, 2018-11-15 15:51
A dataset is a structured version of a source where each field has been processed and serialized according to its type. The possible field types are numeric, categorical, text, date-time, or items. For each field, you can also get the number of errors that were encountered processing it. Errors are mostly missing values or values that do not match with the type assigned to the column.
When you create a new dataset, histograms of the field values are created for the categorical and numeric fields. In addition, for the numeric fields, a collection of statistics about the field distribution such as minimum, maximum, sum, and sum of squares are also computed.
For date-time fields, BigML attempts to parse the format and automatically generate the related subfields (year, month, day, and so on) present in the format.
For items fields which have many different categorical values per instance separated by non-alphanumeric characters, BigML tries to automatically detect which is the best separator for your items.
Finally, for text fields, BigML handles plain text fields with some light-weight natural language processing; BigML separates the field into words using punctuation and whitespace, attempts to detect the language, groups word forms together using word stemming, and eliminates words that are too common or too rare to be useful. We are then left with somewhere between a few dozen and a few hundred interesting words per text field, the occurrences of which can be features in a model.
BigML.io allows you to create, retrieve, update, delete your dataset. You can also list all of your datasets.
Jump to:
- Dataset Base URL
- Creating a Dataset
- Dataset Arguments
- Filtering Rows
- Retrieving a Dataset
- Dataset Properties
- Filtering and Paginating Fields from a Dataset
- Updating a Dataset
- Deleting a Dataset
- Listing Datasets
- Multi-Datasets
- Resources Accepting Multi-Datasets Input
- Creating a Dataset using SQL
- Transformations
- Cloning a Dataset
- Sampling a Dataset
- Filtering a Dataset
- Extending a Dataset
- Filtering the New Fields Output
- Discretization of a Continuous Field
- Outlier Elimination
- Lisp and JSON Syntaxes
- Final Remarks
Dataset Base URL
You can use the following base URL to create, retrieve, update, and delete datasets. https://bigml.io/datasetAll requests to manage your datasets must use HTTPS and be authenticated using your username and API key to verify your identity. See this section for more details.
Creating a Dataset
To create a new dataset, you need to POST to the dataset base URL an object containing at least the source/id that you want to use to create the dataset. The content-type must always be "application/json".
You can easily create a new dataset using curl as follows. All you need is a valid source/id and your authentication variable set up as shown above.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"source": "source/50a4527b3c1920186d000041"}' BigML.io will return the newly created dataset if the request succeeded.
{
"category": 0,
"code": 201,
"columns": 0,
"created": "2012-11-15T02:29:09.293711",
"credits": 0.00439453125,
"description": "",
"excluded_fields": [],
"fields": {},
"fields_meta": {
"count": 0,
"limit": 200,
"offset": 0,
"total": 0
},
"input_fields": [],
"locale": "en-US",
"name": "iris' dataset",
"number_of_evaluations": 0,
"number_of_models": 0,
"number_of_predictions": 0,
"price": 0.0,
"private": true,
"project": null,
"resource": "dataset/52b9359a3c19205ff100002a",
"rows": 0,
"size": 4608,
"source": "source/50a4527b3c1920186d000041",
"source_status": true,
"status": {
"code": 1,
"message": "The dataset is being processed and will be created soon"
},
"tags": [],
"updated": "2012-11-15T02:29:09.293733",
"views": 0
}
Dataset Arguments
By default, the dataset will include all fields in the corresponding source; but this behaviour can be fine-tuned via the input_fields and excluded_fields lists of identifiers. The former specifies the list of fields to be included in the dataset, and defaults to all fields in the source when empty. To specify excluded fields, you can use excluded_fields: identifiers in that list are removed from the list constructed using input_fields".
See below the full list of arguments that you can POST to create a dataset.
| Argument | Type | Description |
|---|---|---|
|
category
optional |
Integer, default is the category of the source |
The category that best describes the dataset. See the category codes for the complete list of categories.
Example: 1 |
|
description
optional |
String |
A description of the dataset up to 8192 characters long.
Example: "This is a description of my new dataset" |
|
excluded_fields
optional |
Array, default is [], an empty list. None of the fields in the source is excluded. |
Specifies the fields that won't be included in the dataset.
Example: |
|
fields
optional |
Object, default is {}, an empty dictionary. That is, no names, labels or descriptions are changed. |
Updates the names, labels, and descriptions of the fields in the dataset with respect to the original names in the source. An entry keyed with the field id generated in the source for each field that you want the name updated.
Example: |
|
input_fields
optional |
Array, default is []. All the fields in the source. |
Specifies the fields to be included in the dataset.
Example: |
|
json_filter
optional |
Array |
A JSON list representing a filter over the rows in the datasource. The first element is an operator and the rest of the elements its arguments. See the section below for more details.
Example: [">", 3.14, ["field", "000002"]] |
|
lisp_filter
optional |
String |
A string representing a Lisp s-expression to filter rows from the datasource.
Example: "(> 3.14 (field 2))" |
|
name
optional |
String, default is source's name |
The name you want to give to the new dataset.
Example: "my new dataset" |
|
objective_field
optional |
Object, default is the last non-auto-generated field in the dataset. |
Specifies the default objective field.
Example: |
| origin | String |
The dataset/id of the gallery dataset to be cloned. The price of the dataset must be 0 to be cloned via API.
Example: "dataset/5b9ab8474e172785e3000003" |
|
project
optional |
String |
The project/id you want the dataset to belong to.
Example: "project/54d98718f0a5ea0b16000000" |
|
refresh_field_types
optional |
Boolean, default is false |
Specifies whether field types need to recomputed or not.
Example: true |
|
refresh_objective
optional |
Boolean, default is false |
Specifies whether the default objective field of the dataset needs to be recomputed or not.
Example: true |
|
refresh_preferred
optional |
Boolean, default is false |
Specifies whether preferred field flags need to recomputed or not.
Example: true |
| shared_hash | String |
The shared hash of the shared dataset to be cloned. The price of the dataset must be 0 to be cloned via API.
Example: "kpY46mNuNVReITw0Z1mAqoQ9ySW" |
|
size
optional |
Integer, default is the source's size |
The number of bytes from the source that you want to use.
Example: 1073741824 |
| source | String |
A valid source/id.
Example: source/4f665b8103ce8920bb000006 |
|
tags
optional |
Array of Strings |
A list of strings that help classify and index your dataset.
Example: ["best customers", "2018"] |
|
term_limit
optional |
Integer |
The maximum total number of terms to be used in text analysis.
Example: 500 |
|
webhook
optional |
Object |
A webhook url and an optional secret phrase. See the Section on Webhooks for more details.
Example: |
You can also use curl to customize a new dataset with a name, and different size, and only a few fields from the original source. For example, to create a new dataset named "my dataset", with only 500 bytes, and with only two fields:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"source": "source/4f665b8103ce8920bb000006", "name": "my dataset", "size": 500, "fields": {"000001": {"name": "width_1"}, "000003": {"name": "width_2"}}}'If you do not specify a name, BigML.io will assign to the new dataset the source's name. If you do not specify a size, BigML.io will use the full the source's size. If you do not specify any fields BigML.io will include all the fields in the source with their corresponding names.
Filtering Rows
The dataset creation request can include an argument, json_filter, specifying a predicate that the input rows from the source have to satisfy in order to be included in the dataset. This predicate is specified as a (possibly nested) JSON list whose first element is an operator and the rest of the elements its arguments. Here's an example of a filter specification to choose only those rows whose field "000002" is less than 3.14:
[">", 3.14, ["field", "000002"]] As you see, the list starts with the operator we want to use, ">", followed by its operands: the number 3.14, and the value of the field with identifier "000002", which is denoted by the operator "field". As another example, this filter:
["=", ["field", "000002"], ["field", "000003"], ["field", "000004"]]selects rows for which the three fields with identifiers "000002", "000003" and "000004" have identical values. Note how you're not limited to two arguments. It's also worth noting that for a filter like that one to be accepted, all three fields must have the same optype (e.g. numeric), otherwise they cannot be compared.
The field operator also accepts as arguments the field's name (as a string) or the row column (as an integer). For instance, if field "000002" had column number 12, and field "000003" was named "Stock prize", our previous query could have been written:
["=", ["field", 12], ["field", "Stock prize"], ["field", "000004"]]If the name is not unique, the first matching field found is picked, consistently over the whole filter formula. If you have duplicated field names, the best thing to do is to use either column numbers or field identifiers in your filters, to avoid ambiguities.
Besides a field's value, one can also ask whether it's missing or not. For instance, to include only those rows for which field "000002" contains a missing token, you would use:
["missing", "000002"]
["and", ["not", ["missing", 12]]
, ["not", ["missing", "Stock prize"]]]
["or", ["=", 3, ["field", "000001"]]
, [">", "1969-07-14T06:10", ["field", "000111"]]
, ["and", ["missing", 23]
, ["=", "Cat", ["field", "000002"]]
, ["<", 2, ["field", "000003"], 4]]] In the examples above, you can also see how dates are allowed and can be compared as numerical values (provided the implied fields are of the correct optype).
Finally, it's also possible to use the arithmetic operators +, -, * and / with numeric fields and constants, as in the following example:
[">", ["/", ["+", ["-", ["field", "000000"]
, 4.4]
, ["field", "000003"]
, ["*", 2
, ["field", "Class"]
, ["field", "000004"]]]
, 3]
, 5.5] These are all the accepted operators:
=, !=, >, >=, <, <=, and, or, not, field, missing, +, -, *, /.To be accepted by the API, the filter must evaluate to a boolean value and contain at least one operator.So, for instance, a constant or a formula evaluating to a number will be rejected.
Since writing and reading the above formula in pure JSON might be a bit involved, you can also send your query to the server as a string representing a Lisp flatine formula using the argument lisp_filter, e.g.
(> (/ (+ (- (field "000000") 4.4)
(field 23)
(* 2 (field "Class") (field "000004")))
3)
5.5)Retrieving a Dataset
Each dataset has a unique identifier in the form "dataset/id" where id is a string of 24 alpha-numeric characters that you can use to retrieve the dataset. Notice that to download the dataset file in the CSV format, you will need to append "/download", and in the Tableau tde format, append "/download?format=tde" to resource id.
To retrieve a dataset with curl:
curl "https://bigml.io/dataset/52b9359a3c19205ff100002a?$BIGML_AUTH"To download the dataset file in the CSV format with curl:
curl "https://bigml.io/dataset/52b9359a3c19205ff100002a/download?$BIGML_AUTH"To download the dataset file in the Tableau tde format with curl:
curl "https://bigml.io/dataset/52b9359a3c19205ff100002a/download?format=tde;$BIGML_AUTH"You can also use your browser to visualize the dataset using the full BigML.io URL or pasting the dataset/id into the BigML.com dashboard.
Dataset Properties
Once a dataset has been successfully created it will have the following properties.
| Property | Type | Description |
|---|---|---|
|
category
filterable, sortable, updatable |
Integer | One of the categories in the table of categories that help classify this resource according to the domain of application. |
| code | Integer | HTTP status code. This will be 201 upon successful creation of the dataset and 200 afterwards. Make sure that you check the code that comes with the status attribute to make sure that the dataset creation has been completed without errors. |
|
columns
filterable, sortable |
Integer | The number of fields in the dataset. |
| correlations | Object |
A dictionary where each entry represents a field (column) in your data with the last calculated correlation/id for it.
Example: |
|
created
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the dataset was created with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
|
credits
filterable, sortable |
Float | The number of credits it cost you to create this dataset. |
|
description
updatable |
String | A text describing the dataset. It can contain restricted markdown to decorate the text. |
| excluded_fields | Array | The list of fields's ids that were excluded to build the model. |
| field_types | Object | A dictionary that informs about the number of fields of each type. It has an entry per each field type (categorical, datetime, numeric, and text), an entry for preferred fields and an entry for the total number of fields. In new datasets, it uses the key effective_fields to inform of the effective number of fields. That is the total number of fields including those created under the hood to support text fields. |
| fields | Object | A dictionary with an entry per field (column) in your data. Each entry includes the column number, the name of the field, the type of the field, and the summary. |
| fields_meta | Object | A dictionary with meta information about the fields dictionary. It specifies the total number of fields, the current offset, and limit, and the number of fields (count) returned. |
| input_fields | Array | The list of input fields' ids used to create the dataset. |
| json_query | Object | A dictionary specifying each of the parts of the executed SQL query that was used to create this dataset. |
| json_query_parsed | Object | The canonical representation of the SQL query as a JSON map. |
|
juxtapose
filterable, sortable |
Boolean | Whether juxtaposition has been performed during creation |
| juxtapose_input_fields | Object | A dictionary keyed by dataset/id and an array of field names and/or ids that specifies the input fields to use for each dataset during merge. |
| locale | String | The source's locale. |
|
name
filterable, sortable, updatable |
String | The name of the dataset as your provided or based on the name of the source by default.) |
|
number_of_anomalies
filterable, sortable |
Integer | The current number of anomalies that use this dataset. |
|
number_of_anomalyscores
filterable, sortable |
Integer | The current number of anomaly scores that use this dataset. |
|
number_of_associations
filterable, sortable |
Integer | The current number of associations that use this dataset. |
|
number_of_associationsets
filterable, sortable |
Integer | The current number of association sets that use this dataset. |
|
number_of_batchanomalyscores
filterable, sortable |
Integer | The current number of batch anomaly scores that use this dataset. |
|
number_of_batchcentroids
filterable, sortable |
Integer | The current number of batch centroids that use this dataset. |
|
number_of_batchpredictions
filterable, sortable |
Integer | The current number of batch predictions that use this dataset. |
|
number_of_batchtopicdistributions
filterable, sortable |
Integer | The current number of batch topic distributions that use this dataset. |
|
number_of_centroids
filterable, sortable |
Integer | The current number of centroids that use this dataset. |
|
number_of_clusters
filterable, sortable |
Integer | The current number of clusters that use this dataset. |
|
number_of_correlations
filterable, sortable |
Integer | The current number of correlations that use this dataset. |
|
number_of_ensembles
filterable, sortable |
Integer | The current number of ensembles that use this dataset. |
|
number_of_evaluations
filterable, sortable |
Integer | The current number of evaluations that use this dataset. |
|
number_of_forecasts
filterable, sortable |
Integer | The current number of forecasts that use this dataset. |
|
number_of_logisticregressions
filterable, sortable |
Integer | The current number of logistic regressions that use this dataset. |
|
number_of_models
filterable, sortable |
Integer | The current number of models that use this dataset. |
|
number_of_optimls
filterable, sortable |
Integer | The current number of OptiMLs that use this dataset. |
|
number_of_predictions
filterable, sortable |
Integer | The current number of predictions that use this dataset. |
|
number_of_statisticaltests
filterable, sortable |
Integer | The current number of statistical tests that use this dataset. |
|
number_of_timeseries
filterable, sortable |
Integer | The current number of time series that use this dataset. |
|
number_of_topicdistributions
filterable, sortable |
Integer | The current number of topic distributions that use this dataset. |
|
number_of_topicmodels
filterable, sortable |
Integer | The current number of topic models that use this dataset. |
|
objective_field
updatable |
Object | The default objective field. |
|
optiml
filterable, sortable |
String | The optiml/id that created this dataset. |
|
optiml_status
filterable, sortable |
Boolean | Whether the OptiML is still available or has been deleted. |
|
origin
filterable, sortable |
String | The dataset/id of the original gallery dataset. |
|
out_of_bag
filterable, sortable |
Boolean | Whether the out-of-bag instances were used to clone the dataset instead of the sampled instances. |
|
price
filterable, sortable, updatable |
Float | The price other users must pay to clone your dataset. |
|
private
filterable, sortable, updatable |
Boolean | Whether the dataset is public or not. |
|
project
filterable, sortable, updatable |
String | The project/id the resource belongs to. |
| range | Array | The range of instances used to clone the dataset. |
|
refresh_field_types
filterable, sortable |
Boolean | Whether the field types of the dataset have been recomputed or not. |
|
refresh_objective
filterable, sortable |
Boolean | Whether the default objective field of the dataset has been recomputed or not. |
|
refresh_preferred
filterable, sortable |
Boolean | Whether the preferred flags of the dataset fields have been recomputed or not. |
|
replacement
filterable, sortable |
Boolean | Whether the instances sampled to clone the dataset were selected using replacement or not. |
| resource | String | The dataset/id. |
|
rows
filterable, sortable |
Integer | The total number of rows in the dataset. |
|
sample_rate
filterable, sortable |
Float | The sample rate used to select instances from the dataset. |
|
seed
filterable, sortable |
String | The string that was used to generate the sample. |
|
shared
filterable, sortable, updatable |
Boolean | Whether the dataset is shared using a private link or not. |
|
shared_clonable
filterable, sortable, updatable |
Boolean | Whether the shared dataset can be cloned or not. |
| shared_hash | String | The hash that gives access to this dataset if it has been shared using a private link. |
| sharing_key | String | The alternative key that gives read access to this dataset. |
|
size
filterable, sortable |
Integer | The number of bytes of the source that were used to create this dataset. |
|
source
filterable, sortable |
String | The source/id that was used to build the dataset. |
|
source_status
filterable, sortable |
Boolean | Whether the source is still available or has been deleted. |
| sql_output_fields | Array of Objects | A list of dictionaries containing some of the properties of the fields generated by the given sql_query or json_query. |
| sql_query | String | The SQL query that was executed to create this dataset. |
| sql_query_parsed | String | The canonical form of query as a SQL prepared statement. |
|
statisticaltest
filterable, sortable |
String | The last statisticaltest/id that was generated for this dataset. |
| status | Object | A description of the status of the dataset. It includes a code, a message, and some extra information. See the table below. |
|
subscription
filterable, sortable |
Boolean | Whether the dataset was created using a subscription plan or not. |
|
tags
filterable, updatable |
Array of Strings | A list of user tags that can help classify and index this resource. |
|
term_limit
filterable, sortable |
Integer | The maximum total number of terms used by all the text fields. |
|
updated
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the dataset was updated with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
| webhook | Object | A webhook url and an optional secret phrase. See the Section on Webhooks for more details. |
Dataset Fields
The property fields is a dictionary keyed by each field's id in the source. Each field's id has as a value an object with the following properties:
Numeric Summary
Numeric summaries come with all the fields described below. If the number of unique values in the data is greater than 32, then 'bins' will be used for the summary. If not, 'counts' will be available.
| Property | Type | Description |
|---|---|---|
| bins | Array | An array that represents an approximate histogram of the distribution. It consists of value pairs, where the first value is the mean of a histogram bin and the second value is the bin population. bins is only available when the number of distinct values is greater than 32. For more information, see our blog post or read this paper. |
| counts | Array | An array of pairs where the first element of each pair is one of the unique values found in the field and the second element is the count. Only available when the number of distinct values is less than or equal to 32. |
| kurtosis | Number | The sample kurtosis. A measure of 'peakiness' or heavy tails in the field's distribution. |
| maximum | Number | The maximum value found in this field. |
| mean | Number | The arithmetic mean of non-missing field values. |
| median | Number | The approximate median of the non-missing values in this field. |
| minimum | Number | The minimum value found in this field. |
| missing_count | Integer | Number of instances missing this field. |
| population | Integer | The number of instances containing data for this field. |
| skewness | Number | The sample skewness. A measure of asymmetry in the field's distribution. |
| standard_deviation | Number | The unbiased sample standard deviation. |
| sum | String | Sum of all values for this field (for mean calculation). |
| sum_squares | String | Sum of squared values (for variance calculation). |
| variance | Number | The unbiased sample variance. |
Categorical Summary
Categorical summaries give you a count per each category and missing count in case any of the instances contain missing values.
Text Summary
Text summaries give statistics about the vocabulary of a text field, and the number of instances containing missing values.
Dataset Status
Before a dataset is successfully created, BigML.io makes sure that it has been uploaded in an understandable format, that the data that it contains is parseable, and that the types for each column in the data can be inferred successfully. The dataset goes through a number of states until all these analyses are completed. Through the status field in the dataset you can determine when the dataset has been fully processed and ready to be used to create a model. These are the fields that a dataset's status has:
| Property | Type | Description |
|---|---|---|
| bytes | Integer | Number of bytes processed so far. |
| code | Integer | A status code that reflects the status of the dataset creation. It can be any of those that are explained here. |
| elapsed | Integer | Number of milliseconds that BigML.io took to process the dataset. |
| field_errors | Object |
Information about ill-formatted fields that includes the total format errors for the field and a sample of the ill-formatted tokens.
Example: |
| message | String | A human readable message explaining the status. |
| row_format_errors | Array | Information about ill-formatted rows. It includes the total row-format errors and a sampling of the ill-formatted rows. |
| serialized_rows | Integer | The number of rows serialized so far. |
Once a dataset has been successfully created, it will look like:
{
"category": 0,
"code": 200,
"columns": 5,
"created": "2012-11-15T02:29:09.293000",
"credits": 0.00439453125,
"description": "",
"excluded_fields": [],
"fields": {
"000000": {
"column_number": 0,
"datatype": "double",
"name": "sepal length",
"optype": "numeric",
"order": 0,
"preferred": true,
"summary": {
"bins": [
[
4.3,
1
],
[
4.425,
4
],
[
4.6,
4
],
[
4.7,
2
],
[
4.8,
5
],
[
4.9,
6
],
[
5,
10
],
[
5.1,
9
],
[
5.2,
4
],
[
5.3,
1
],
[
5.4,
6
],
[
5.5,
7
],
[
5.6,
6
],
[
5.7,
8
],
[
5.8,
7
],
[
5.9,
3
],
[
6,
6
],
[
6.1,
6
],
[
6.2,
4
],
[
6.3,
9
],
[
6.44167,
12
],
[
6.6,
2
],
[
6.7,
8
],
[
6.8,
3
],
[
6.92,
5
],
[
7.1,
1
],
[
7.2,
3
],
[
7.3,
1
],
[
7.4,
1
],
[
7.6,
1
],
[
7.7,
4
],
[
7.9,
1
]
],
"maximum": 7.9,
"mean": 5.84333,
"median": 5.77889,
"minimum": 4.3,
"missing_count": 0,
"population": 150,
"splits": [
4.51526,
4.67252,
4.81113,
4.89582,
4.96139,
5.01131,
5.05992,
5.11148,
5.18177,
5.35681,
5.44129,
5.5108,
5.58255,
5.65532,
5.71658,
5.77889,
5.85381,
5.97078,
6.05104,
6.13074,
6.23023,
6.29578,
6.35078,
6.41459,
6.49383,
6.63013,
6.70719,
6.79218,
6.92597,
7.20423,
7.64746
],
"standard_deviation": 0.82807,
"sum": 876.5,
"sum_squares": 5223.85,
"variance": 0.68569
}
},
"000001": {
"column_number": 1,
"datatype": "double",
"name": "sepal width",
"optype": "numeric",
"order": 1,
"preferred": true,
"summary": {
"counts": [
[
2,
1
],
[
2.2,
3
],
[
2.3,
4
],
[
2.4,
3
],
[
2.5,
8
],
[
2.6,
5
],
[
2.7,
9
],
[
2.8,
14
],
[
2.9,
10
],
[
3,
26
],
[
3.1,
11
],
[
3.2,
13
],
[
3.3,
6
],
[
3.4,
12
],
[
3.5,
6
],
[
3.6,
4
],
[
3.7,
3
],
[
3.8,
6
],
[
3.9,
2
],
[
4,
1
],
[
4.1,
1
],
[
4.2,
1
],
[
4.4,
1
]
],
"maximum": 4.4,
"mean": 3.05733,
"median": 3.02044,
"minimum": 2,
"missing_count": 0,
"population": 150,
"standard_deviation": 0.43587,
"sum": 458.6,
"sum_squares": 1430.4,
"variance": 0.18998
}
},
"000002": {
"column_number": 2,
"datatype": "double",
"name": "petal length",
"optype": "numeric",
"order": 2,
"preferred": true,
"summary": {
"bins": [
[
1,
1
],
[
1.1,
1
],
[
1.2,
2
],
[
1.3,
7
],
[
1.4,
13
],
[
1.5,
13
],
[
1.63636,
11
],
[
1.9,
2
],
[
3,
1
],
[
3.3,
2
],
[
3.5,
2
],
[
3.6,
1
],
[
3.75,
2
],
[
3.9,
3
],
[
4.0375,
8
],
[
4.23333,
6
],
[
4.46667,
12
],
[
4.6,
3
],
[
4.74444,
9
],
[
4.94444,
9
],
[
5.1,
8
],
[
5.25,
4
],
[
5.46,
5
],
[
5.6,
6
],
[
5.75,
6
],
[
5.95,
4
],
[
6.1,
3
],
[
6.3,
1
],
[
6.4,
1
],
[
6.6,
1
],
[
6.7,
2
],
[
6.9,
1
]
],
"maximum": 6.9,
"mean": 3.758,
"median": 4.34142,
"minimum": 1,
"missing_count": 0,
"population": 150,
"splits": [
1.25138,
1.32426,
1.37171,
1.40962,
1.44567,
1.48173,
1.51859,
1.56301,
1.6255,
1.74645,
3.23033,
3.675,
3.94203,
4.0469,
4.18243,
4.34142,
4.45309,
4.51823,
4.61771,
4.72566,
4.83445,
4.93363,
5.03807,
5.1064,
5.20938,
5.43979,
5.5744,
5.6646,
5.81496,
6.02913,
6.38125
],
"standard_deviation": 1.7653,
"sum": 563.7,
"sum_squares": 2582.71,
"variance": 3.11628
}
},
"000003": {
"column_number": 3,
"datatype": "double",
"name": "petal width",
"optype": "numeric",
"order": 3,
"preferred": true,
"summary": {
"counts": [
[
0.1,
5
],
[
0.2,
29
],
[
0.3,
7
],
[
0.4,
7
],
[
0.5,
1
],
[
0.6,
1
],
[
1,
7
],
[
1.1,
3
],
[
1.2,
5
],
[
1.3,
13
],
[
1.4,
8
],
[
1.5,
12
],
[
1.6,
4
],
[
1.7,
2
],
[
1.8,
12
],
[
1.9,
5
],
[
2,
6
],
[
2.1,
6
],
[
2.2,
3
],
[
2.3,
8
],
[
2.4,
3
],
[
2.5,
3
]
],
"maximum": 2.5,
"mean": 1.19933,
"median": 1.32848,
"minimum": 0.1,
"missing_count": 0,
"population": 150,
"standard_deviation": 0.76224,
"sum": 179.9,
"sum_squares": 302.33,
"variance": 0.58101
}
},
"000004": {
"column_number": 4,
"datatype": "string",
"name": "species",
"optype": "categorical",
"order": 4,
"preferred": true,
"summary": {
"categories": [
[
"Iris-versicolor",
50
],
[
"Iris-setosa",
50
],
[
"Iris-virginica",
50
]
],
"missing_count": 0
}
}
},
"fields_meta": {
"count": 5,
"limit": 200,
"offset": 0,
"total": 5
},
"input_fields": [
"000000",
"000001",
"000002",
"000003",
"000004"
],
"locale": "en_US",
"name": "iris' dataset",
"number_of_evaluations": 0,
"number_of_models": 0,
"number_of_predictions": 0,
"price": 0.0,
"private": true,
"project": null,
"resource": "dataset/52b9359a3c19205ff100002a",
"rows": 150,
"size": 4608,
"source": "source/50a4527b3c1920186d000041",
"source_status": true,
"status": {
"bytes": 4608,
"code": 5,
"elapsed": 163,
"field_errors": [],
"message": "The dataset has been created",
"row_format_errors": [],
"serialized_rows": 150
},
"tags": [],
"updated": "2012-11-15T02:29:10.537000",
"views": 0
}
Filtering and Paginating Fields from a Dataset
A dataset might be composed of hundreds or even thousands of fields. Thus when retrieving a dataset, it's possible to specify that only a subset of fields be retrieved, by using any combination of the following parameters in the query string (unrecognized parameters are ignored):
Since fields is a map and therefore not ordered, the returned fields contain an additional key, order, whose integer (increasing) value gives you their ordering. In all other respects, the source is the same as the one you would get without any filtering parameter above.
The fields_meta field can help you paginate fields. Its structure is as follows:Updating a Dataset
To update a dataset, you need to PUT an object containing the fields that you want to update to the dataset' s base URL. The content-type must always be: "application/json". If the request succeeds, BigML.io will return with an HTTP 202 response with the updated dataset.
For example, to update a dataset with a new name you can use curl like this:
curl "https://bigml.io/dataset/52b9359a3c19205ff100002a?$BIGML_AUTH" \
-X PUT \
-d '{"name": "a new name"}' \
-H 'content-type: application/json'Deleting a Dataset
To delete a dataset, you need to issue a HTTP DELETE request to the dataset/id to be deleted.
Using curl you can do something like this to delete a dataset:
curl -X DELETE "https://bigml.io/dataset/52b9359a3c19205ff100002a?$BIGML_AUTH"If the request succeeds you will not see anything on the command line unless you executed the command in verbose mode. Successful DELETEs will return "204 no content" responses with no body.
Once you delete a dataset, it is permanently deleted. That is, a delete request cannot be undone. If you try to delete a dataset a second time, or a dataset that does not exist, you will receive a "404 not found" response.
However, if you try to delete a dataset that is being used at the moment, then BigML.io will not accept the request and will respond with a "400 bad request" response.
See this section for more details.Listing Datasets
To list all the datasets, you can use the dataset base URL. By default, only the 20 most recent datasets will be returned. You can see below how to change this number using the limit parameter.
You can get your list of datasets directly in your browser using your own username and API key with the following links.
https://bigml.io/dataset?$BIGML_AUTHMulti-Datasets
BigML.io now allows you to create a new dataset merging multiple datasets. This functionaliy can be very useful when you use multiple sources of data and in online scenarios as well. Imagine, for example, that you collect data in a hourly basis and want to create a dataset aggregrating data collected over the whole day. So you only need to send the new generated data each hour to BigML, create a source and a dataset for each one and then merge all the individual datasets into one at the end of the day.
We usually call datasets created in this way multi-datasets. BigML.io allows you to aggregrate up to 32 datasets in the same API request. You can merge multi-datasets too so basically you can grow a dataset as much as you want.
To create a multi dataset, you can specify a list of dataset identifiers as input using the argument origin_datasets. The example below will construct a new dataset that is the concatenation of three other datasets.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc7fd03c1920e4a3000016",
"dataset/52bc80233c1920e4a300001a"]}' By convention, the first dataset defines the final dataset fields. However, there can be cases where each dataset might come from a different source and therefore have different field ids. In these cases, you might need to use a fields_maps argument to match each field in a dataset to the fields of the first dataset.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc7fd03c1920e4a3000016",
"dataset/52bc80233c1920e4a300001a",
"dataset/52bc851b3c1920e4a3000022"],
"fields_maps": {
"dataset/52bc7fd03c1920e4a3000016": {
"000000":"000023",
"000001":"000024",
"000002":"00003a"},
"dataset/52bc80233c1920e4a300001a": {
"000000":"000023",
"000001":"000004",
"000002":"00000f"}}}' For instance, in the request above, we use four datasets as input. The first one would define the final dataset fields. For instance, let's say that the dataset "dataset/52bc7fc83c1920e4a3000012" in this example has three fields with identifiers "000001", "000002" and "000003". Those will be the default resulting fields, together with their datatypes and so on. Then we need to specify, for each of the remaining datasets in the list, a mapping from the "standard" fields to those in the corresponding dataset. In our example, we're saying that the fields of the second dataset to be used during the concatenation are "000023", "000024" and "00003a", which correspond to the final fields having them as keys. In the case of the third dataset, the fields used will be "000023", "000004" and "00000f". For the last one, since there's no entry in fields_maps, we'll try to use the same identifiers as those of the first dataset.
The optypes of the paired fields should match, and for the case of categorical fields, be a proper subset. If a final field has optype text, however, all values are converted to strings.
BigML.io also allows you to sample each dataset individually before merging it. You can specify the sample options for each dataset using the arguments sample_rates, replacements, seeds, and out_of_bags. All are dictionaries that must be keyed using the dataset/id of the dataset you want to specify parameters for. The next request will create a multi-dataset sampling the two input datasets differently.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc851b3c1920e4a3000022"],
"sample_rates": {
"dataset/52bc7fc83c1920e4a3000012": 0.5,
"dataset/52bc851b3c1920e4a3000022": 0.8},
"replacements": {
"dataset/52bc7fc83c1920e4a3000012": false,
"dataset/52bc851b3c1920e4a3000022": true}
}'BigML.io also allows you to create a new dataset merging multiple datasets using juxtaposition instead of concatenating the datasets passed in the argument origin_datasets. In its simplest form, a juxtaposition request would look like this:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc7fd03c1920e4a3000016",
"dataset/52bc80233c1920e4a300001a"],
"juxtapose": true}' In the example above we are asking for the generation of a new dataset with each row constructed by concatenating the three rows of each origin dataset put side by side. The new dataset will thus contain as many rows as the shorter input dataset and as many fields as the sum of the number of fields of the input datasets.
Unless otherwise specified, all fields of each of the datasets in origin_datasets is used in the juxtaposition. If you want to use a subset of any of them, specify it using juxtapose_input_fields. This creation request field must be an object where each entry specifies the input fields to use for the corresponding dataset. The values must be a list of fields names or identifiers. For instance:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc7fd03c1920e4a3000016",
"dataset/52bc80233c1920e4a300001a"],
"juxtapose": true,
"juxtapose_input_fields": {
"dataset/52bc7fc83c1920e4a3000012": ["000001", "species"],
"dataset/52bc80233c1920e4a300001a": ["age", "000002", "000003"]}}' It will juxtapose two fields of the first dataset, all the fields of the second dataset, and three fields of the last dataset. We also show in the example how fields can be identified by either id or name.
This is the list of all the arguments that you can use to create a multi dataset.| Argument | Type | Description |
|---|---|---|
|
fields_maps
optional |
Object |
A dictionary keyed by dataset/id and object values. Each entry maps fields in the first dataset to fields in the dataset referenced by the key.
Example: |
|
json_query
optional |
Object | A dictionary specifying each of the parts of the executed SQL query separately. See the Section on Creating a Dataset using SQL for more details. |
|
juxtapose
optional |
Boolean, default is false |
Whether juxtaposition should be performed on multi-dataset merging.
Example: true |
|
juxtapose_input_fields
optional |
Object |
A dictionary keyed by dataset/id and an array of field names and/or ids that specifies the input fields to use for each dataset during merge.
Example: |
|
origin_dataset_names
optional |
Object | A dictionary keyed by dataset/id and value a string that represents the name to be used as its table name in the SQL query to be performed for the dataset. See the Section on Creating a Dataset using SQL for more details. |
|
out_of_bags
optional |
Object |
A dictionary keyed by dataset/id and boolean values. Setting this parameter to true for a dataset will return a dataset containing sequence of the out-of-bag instances instead of the sampled instances. See the Section on Sampling for more details.
Example: |
|
replacements
optional |
Object |
A dictionary keyed by dataset/id and boolean values indicating whether sampling should be performed with or without replacement. See the Section on Sampling for more details.
Example: |
|
sample_rates
optional |
Object |
A dictionary keyed by dataset/id and float values. Each value is a number between 0 and 1 specifying the sample rate for the dataset. See the Section on Sampling for more details.
Example: |
|
seeds
optional |
Object |
A dictionary keyed by dataset/idand string values indicating the seed to be used for each dataset to generate deterministic samples. See the Section on Sampling for more details.
Example: |
|
sql_output_fields
optional |
Array of Objects | A list of dictionaries containing some of the properties of the fields generated by the given sql_query or json_query. See the Section on Creating a Dataset using SQL for more details. |
|
sql_query
optional |
String | The SQL query to be executed.See the Section on Creating a Dataset using SQL for more details. |
- Sample each individual dataset according to the specifications provided in the arguments sample_rates, replacements, seeds, and out_of_bags.
- Merge all the datasets together using the fields_maps argument to match fields in case they come from different sources (i.e., have different field ids).
- When juxtapose is true, all argument in the table above except juxtapose_input_fields are ignore, however, the rules below still apply.
- Sample the merged dataset like in the case of a regular datasaset sampling using the the arguments sample_rate, replacement, seed,out_of_bag.
- Filter the sampled dataset using input_fields, excluded_fields, and either a json_filter or lisp_filter.
- Extend the dataset with new fields according to the specifications provided in the new_fields argument.
- Filter the output of the new fields using either an output_json_filter or output_lisp_filter.
Resources Accepting Multi-Datasets Input
You can also create a model using multiple datasets as input at once. That is, without merging all the datasets together into a new dataset first. The same applies to correlations, statistical tests, ensembles, clusters, anomaly detectors, and evaluations. All the multi-dataset arguments above can be used. You just need to use the datasets argument instead of the regular dataset.See examples below to create a multi-dataset model, a multi-dataset ensemble, and a multi-dataset evaluation.
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc851b3c1920e4a3000022"],
"sample_rates": {
"dataset/52bc7fc83c1920e4a3000012": 0.5,
"dataset/52bc851b3c1920e4a3000022": 0.8},
"replacements": {
"dataset/52bc7fc83c1920e4a3000012": false,
"dataset/52bc851b3c1920e4a3000022": true}
}'
curl "https://bigml.io/ensemble?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc851b3c1920e4a3000022"],
"sample_rates": {
"dataset/52bc7fc83c1920e4a3000012": 0.5,
"dataset/52bc851b3c1920e4a3000022": 0.8},
"replacements": {
"dataset/52bc7fc83c1920e4a3000012": false,
"dataset/52bc851b3c1920e4a3000022": true}
}'
curl "https://bigml.io/evaluation?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"model": "model/52bcb43e3c1920e4a3000026",
"datasets": [
"dataset/52bc7fc83c1920e4a3000012",
"dataset/52bc851b3c1920e4a3000022"],
"out_of_bags": {
"dataset/52bc7fc83c1920e4a3000012": true,
"dataset/52bc851b3c1920e4a3000022": true},
"sample_rates": {
"dataset/52bc7fc83c1920e4a3000012": 0.5,
"dataset/52bc851b3c1920e4a3000022": 0.8},
"replacements": {
"dataset/52bc7fc83c1920e4a3000012": false,
"dataset/52bc851b3c1920e4a3000022": true}
}'Creating a Dataset using SQL
BigML.io now allows you to create a new dataset by performing an SQL-style query over a list of input datasets, which are treated as SQL tables. To that end, the POST request JSON should contain the following fields:
- origin_datasets: a list of identifiers for the input datasets that are going to be used as the input tables of the query, this is identical to the field used to specify multiple datasets during merging datasets.
- origin_dataset_names: a dictionary keyed by dataset/id and value a string that represents the name to be used as its table name in the SQL query to be performed for the dataset. This will typically be a list of sort names so that the SQL query is readable. (i.e., you want to write "SELECT d0.field1" rather than "SELECT a_long_dataset_name.field1".)
- sql_query, a string with the SQL query to be executed or json_query, a map specifying each of the parts of the SQL query separately.
- sql_output_fields, a list of dictionaries containing some of the properties of the fields generated by the given sql_query or json_query.
The platform will parse the query, converting if needed its field names to identifiers and return back (for informational purposes) two additional fields, namely sql_query_parsed and json_query_parsed. The first is the canonical form of query as a SQL prepared statement (i.e., a list with a string that can contain wild-cards and, if needed, some arguments as in ["SELECT * FROM A WHERE A.000000 > ?", 2]), and json_query is the canonical representation of query as a JSON map.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/5bab8d6b1f386f7c20000000",
"dataset/5bab8d6e1f386f7c20000003"
],
"origin_dataset_names": {
"dataset/5bab8d6b1f386f7c20000000": "A",
"dataset/5bab8d6e1f386f7c20000003": "B"
},
"sql_query": "select A.`000000` as x, A.`00000a` as z, A.`00000c` from A, B where A.id = B.id",
"sql_output_fields": [
{
"column": 0,
"name": "name, a text",
"optype": "text",
"term_analysis": {
"enabled": true,
"case_sensitive": true
}
},
{
"column": 1,
"optype": "items",
"item_analysis": {
"separator": ";"
}
}
]
}' A SELECT specification can be provided either as a SQL string or as a map possibly containing the following keys:
-
SELECT: A list of strings,
each one specifying one of the new fields in the generated dataset.
This corresponds to the "selected columns" part of a "SELECT FROM ..." SQL statement,
and will use the names in origin_dataset_names
to refer to input datasets as SQL tables.
Each table has as one column per dataset field, and its canonical name is the field identifier;
but for convenience one can refers to input columns using field names
and BigML will translate them automatically to identifiers.
For instance, say we have origin_dataset_names:
{"dataset/5bab8d6b1f386f7c20000000": "d0"}, i.e. one input dataset with,
say, fields 000000, 000001 and 000002 named field1, field2 and field3.
One could select only the first field of the first dataset either via
"SELECT d0.000000" or via "SELECT d0.field1", or using the maps:
{"select": ["d0.`000000`"], ...}
{"select": ["d0.field1"], ...}
or the second and first columns of the dataset with, for instance, "SELECT d0.000001, d0.field1", with the map form:{"select": ["d0.`000001`", "d0.field1"], ...}
In an SQL query specified as string, you can name the output columns of your query using "AS", for instance: "SELECT d0.field2 AS age" will pick the third column of the dataset and the field of the generated dataset will be named age. When the query is specified using a JSON dictionary, the corresponding element in the select list will be a pair, with the first element the left hand operator of "AS" and the second element the right hand one. So the previous query would be translated as:{"select": [["d0.field2", "age"]], ...}
It is also possible to specify SQL function calls in a select list element, using prefix notation for the operators. - DISTINCT: A boolean flag that corresponds to the distinct SQL keyword when set to true. Defaults to false.
- LIMIT: An integer with the maximum number of rows to select, as in the SQL string "SELECT * LIMIT 3".
- OFFSET: The offset of the selected rows, as an integer. Corresponds to the offset SQL keyword.
-
WHERE: A JSON rendition, as a list, of a SQL where clause.
The first element in the list is the SQL operand to apply
(one of and, or, count,
avg, sum, min,
max, =, <>,
!=, >, >=,
<, <=, and between),
and the rest are its operands (possibly including nested operators).
So for instance, the SQL clause "d.f0 < 3 and e.f1 = e.f2" is written as
["and", ["<", "d.f0", 3], ["=", "e.f1", "e.f2"]]
- HAVING: A translation of a SQL having clause using prefix notation, as in where.
- GROUP_BY: A list of field identifiers (as in select) to perform a SQL group by operation.
-
ORDER_BY: A list of field identifiers to perform a SQL order by operation.
Each element in the list can be either a field identifier string,
or a pair of a field identifier and either ASC or
DESC to denote ascending or descending ordering.
["t.field1", ["t.field2", "DESC"], ["e.field0" "ASC"]]
-
JOIN, LEFT_JOIN, RIGHT_JOIN, FULL_JOIN:
An SQL JOIN clause is used to combine rows from two or more datasets (tables).
The specification consists of a list starting with the name of the dataset to join on,
followed by the operation that one writes in the SQL ON specification.
Thus, for instance, the SQL string "JOIN foo ON foo.id = bar.id" would be translate to the JSON specification
{"join": ["foo", ["=", "foo.id", "bar.id"]]}
and likewise for the other joins.
Here's an example of a complicated query combining most of the elements above:
{
"select": ["f.*", "b.baz", "c.quux", ["b.bla", "bla-bla"], ["now"]],
"distinct": true,
"having": ["<", 0, "f.e"],
"where": ["or", ["and", ["=", "f.a", "bort"], ["!=", "b.baz", "param1"]],
["<", 1, 2, 3],
["in", "f.e", [1, 19, 3]],
["between", "f.e", 10, 20]],
"limit": 50,
"group_by": ["f.a"],
"offset": 10,
"join": ["draq", ["=", "f.b", "draq.x"]],
"left_join": ["clod", ["=", "f.a", "clod.d"]],
"order_by": [["b.baz", "desc"], "c.quux", "f.a"]
}which would correspond to the SQL query string:
SELECT DISTINCT f.*, b.baz, c.quux, b.bla AS bla_bla, now()
INNER JOIN draq ON f.b = draq.x
LEFT JOIN clod c ON f.a = c.d
WHERE ((f.a = "bort" AND b.baz <> "param1")
OR (1 < 2 AND 2 < 3)
OR (f.e in (1, 10, 3))
OR f.e BETWEEN 10 AND 20)
GROUP BY f.a
HAVING 0 < f.e
ORDER BY b.baz DESC, c.quux, f.a
LIMIT 50
OFFSET 10Although the above examples contain field names, you can also use the field IDs in a SQL query as long as you set them with back quotes, e.g., “SELECT `100006`, sum(`000001`) AS sum_field1, FROM DS GROUP BY `100006`”
The user can submit either form as her query. If she uses the latter, as a string, BigML will parse it to a standard map format as the former, discarding non-supported SQL constructs appearing in the query string.
Note that, as is conventional in SQL, we mix freely upper and lowercase keywords in the above examples for a sql_query value. BigML accepts both cases although the recommended style is to not mix them in a single request. The keywords, however, in json_query must be in lowercase.
Next, we'll list all the arguments that can be used to fine-tuning the properties of the SQL-generated fields.
| Argument | Type | Description |
|---|---|---|
| column | Integer |
Column number denoting the output field. Zero-based numbering.
Example: 1 |
|
item_analysis
optional |
Object |
Set of parameters to activate item analysis for the dataset. See the Section on Item Analysis for more details.
Example: |
|
name
optional |
String |
Name of the new field.
Example: "Price" |
|
optype
optional |
String |
Optype of the new field. Available optypes are "numeric", "categorical", "text", "datetime", and "items".
Example: "text" |
|
refresh_field_type
optional |
Boolean, default is false |
Whether the optype of the field needs to recomputed or not.
Example: true |
|
refresh_preferred
optional |
Boolean, default is false |
Whether the preferred flag of the field needs to recomputed or not.
Example: true |
|
term_analysis
optional |
Object |
Set of parameters to activate text analysis for the dataset. See the Section on Term Analysis for more details.
Example: |
Column Concatenation Example
Say you have 2 datasets and want to join them using the field named "id". A query request would look like this:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/5bab8d6b1f386f7c20000000",
"dataset/5bab8d6e1f386f7c20000003"],
"origin_dataset_names": {
"dataset/5bab8d6b1f386f7c20000000": "A",
"dataset/5bab8d6e1f386f7c20000003": "B"},
"sql_query": "select * from A join B on A.id=B.id"
}'or, using the JSON dictionary form of the query:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_datasets": [
"dataset/5bab8d6b1f386f7c20000000",
"dataset/5bab8d6e1f386f7c20000003"],
"origin_dataset_names": {
"dataset/5bab8d6b1f386f7c20000000": "A",
"dataset/5bab8d6e1f386f7c20000003": "B"},
"json_query": {
"select": ["*"],
"from": ["A"],
"join": ["B", ["=", "A.id", "B.id"]]}
}' Transformations
Once you have created a dataset, BigML.io allows you to derive new datasets from it, sampling, filtering, adding new fields, or concatenating it to other datasets. We apply the term dataset transformations to the set of operations to create new modified versions of your original dataset or just transformations to abbreviate.
We use the term:- Cloning for the general operation of generating a new dataset.
- Sampling when the original dataset is sampled.
- Filtering when the original dataset is filtered.
- Extending when new fields are generated.
- Merging when a multi-dataset is created.
Keep in mind that you can sample, filter and extend a dataset all at once in only one API request.
So let's start with the most basic transformation: cloning a dataset.
Cloning a Dataset
To clone a dataset you just need to use the origin_dataset argument to send the dataset/id of the dataset that you want to clone. For example:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b8fdff3c19205ff100001e"}' | Argument | Type | Description |
|---|---|---|
|
category
optional |
Integer |
The category that best describes the dataset. See the category codes for the complete list of categories.
Example: 1 |
|
fields
optional |
Object |
Updates the names, labels, and descriptions of the fields in the new dataset. An entry keyed with the field id of the original dataset for each field that will be updated.
Example: |
|
name
optional |
String |
The name you want to give to the new dataset.
Example: "my new dataset" |
| origin_dataset | String |
The dataset/id of the dataset to be cloned.
Example: "dataset/52694b59035d0737c201ac68" |
Sampling a Dataset
It is also possible to provide a sampling specification to be used when cloning the dataset. The sample will be applied to the origin_dataset rows. For example:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b8fdff3c19205ff100001e",
"sample_rate": 0.8,
"replacement": true,
"seed": "myseed"}' | Argument | Type | Description |
|---|---|---|
|
out_of_bag
optional |
Boolean, default is false |
Setting this parameter to true will return a dataset containing a sequence of the out-of-bag instances instead of the sampled instances. See the Section on Sampling for more details.
Example: true |
|
replacement
optional |
Boolean, default is false |
Whether sampling should be performed with or without replacement. See the Section on Sampling for more details.
Example: true |
|
sample_rate
optional |
Float, default is 1.0 |
A real number between 0 and 1 specifying the sample rate. See the Section on Sampling for more details.
Example: 0.5 |
|
seed
optional |
String |
A string to be hashed to generate deterministic samples. See the Section on Sampling for more details.
Example: "MySample" |
Filtering a Dataset
A dataset can be filtered in different ways:- Excluding a few fields using the excluded_fields argument.
- Selecting only a few fields using the input_fields argument.
- Filtering rows using a json_filter or lisp_filter similarly to the way you can filter a source.
- Specifying a range of rows.
As illustrated in the following example, it's possible to provide a list of input fields, selecting the fields from the filtered input dataset that will be created. Filtering happens before field picking and, therefore, the row filter can use fields that won't end up in the cloned dataset.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b8fdff3c19205ff100001e",
"input_fields": ["000000", "000001", "000003"],
"json_filter": [">", 3.14, ["field", "000002"]],
"range": [50, 100]}' | Argument | Type | Description |
|---|---|---|
|
excluded_fields
optional |
Array |
Specifies the fields that won't be included in the new dataset.
Example: |
|
input_fields
optional |
Array |
Specifies the fields to be included in the dataset.
Example: |
|
json_filter
optional |
Array |
A JSON list representing a filter over the rows in the origin dataset. The first element is an operator and the rest of the elements its arguments. See the Section on filtering sources for more details.
Example: [">", 3.14, ["field", "000002"]] |
|
lisp_filter
optional |
String |
A string representing a Lisp s-expression to filter rows from the origin dataset.
Example: "(> 3.14 (field 2))" |
|
range
optional |
Array |
The range of successive instances to create the new dataset.
Example: [100, 200] |
Extending a Dataset
You can clone a dataset and extend it with brand new fields using the new_fields argument. Each new field is created using a Flatline formula and optionally a name, label, and description.
A Flatline formula is a lisp-like expresion that allows you to make references and process columns and rows of the origin dataset. See the full Flatline reference here. Let's see a first example that clones a dataset and adds a new field named "Celsius" to it using a formula that converts the values from the "Fahrenheit" field to Celsius.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb64713c192015e3000010",
"new_fields": [{
"field": "(/ (* 5 (- (f Fahrenheit) 32)) 9)",
"name": "Celsius"}]}'
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb64713c192015e3000010",
"all_fields": false,
"new_fields": [{
"fields": "(fields 0 1)",
"names": ["Day", "Temperature"]}]}'
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb64713c192015e3000010",
"new_fields": [
{"field": "(avg (window Fahrenheit -6 0))",
"name": "Weekly AVG",
"label":"Weekly Average",
"description": "Temperature average over the last seven days"},
{"fields": "(list (f 0 -1) (f 0 1))",
"names": ["Yesterday", "Tomorrow"],
"labels": ["Yesterday prediction", "Tomorrow prediction"],
"descriptions": ["Prediction for the previous day", "Prediction for the next day"]}]}'Filtering the New Fields Output
The generation of new fields works by traversing the input dataset row by row and applying the Flatline formula of each new field to each row in turn. The list of values generated from each input row that way constitutes an output row of the generated dataset.
It is possible to limit the number of input rows that the generator sees by means of filters and/or sample specifications, for example:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb2c263c192015e3000004",
"lisp_filter": "(not (= 0 (f 000001)))",
"new_fields": [
{"field": "(/ 1 (f 000001))",
"name": "Inverse value"}]}'And, as an additional convenience, it is also possible to specify either a output_lisp_filter or a output_json_filter, that is, a Flatline row filter that will act on the generated rows, instead of on the input data:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb2c263c192015e3000004",
"lisp_filter": "(not (= 0 (f 000001)))",
"new_fields": [
{"field": "(/ 1 (f 000001))",
"name": "Inverse value"}],
"output_lisp_filter": "(< 0.25 (f \"Inverse value\"))"}'You can also skip any number of rows in the input, starting the generation at an offset given by row_offset, and traverse the input rows by any step specified by row_step. For instance, the following request will generate a dataset whose rows are created by putting together every three consecutive values of the input field "Price":
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b7f0ba3c19208c13000131",
"row_offset": 2,
"row_step": 3,
"new_fields": [
{"fields": "(window \"Price\" -2 0)",
"names": ["Price-2", "Price-1", "Price"]}]}'With the specification above, the new field will start with the third row in the input dataset, generate an output row (which uses values from the current input row as well as from the two previous ones), skip to the 6th input row, generate a new output, and so on and so forth.
Next, we'll list all the arguments that can be used to extend a dataset.
| Argument | Type | Description |
|---|---|---|
|
all_but
optional |
Array |
Specifies the fields to be included in the dataset.
Example: ["000001", "000003"] |
|
all_fields
optional |
Boolean |
Whether all fields should be included in the new dataset or not.
Example: false |
|
new_fields
optional |
Array |
Specifies the new fields to be included in the dataset. See the table below for more details.
Example: [{"field": "(log10 (field "000001"))", "name": "log"}] |
|
output_json_filter
optional |
Array |
A JSON list representing a filter over the rows of the dataset once the new fields have been generated. The first element is an operator and the rest of the elements its arguments. See the Section on filtering rows for more details.
Example: [">", 3.14, ["field", "000002"]] |
|
output_lisp_filter
optional |
String |
A string representing a Lisp s-expression to filter rows after the new fields have been generated.
Example: "(> 3.14 (field 2))" |
|
row_offset
optional |
Array |
The initial number of rows to skip from from the input dataset before start processing rows.
Example: 100 |
|
row_step
optional |
Array |
The number of rows to skip in every step.
Example: 5 |
| Argument | Type | Description |
|---|---|---|
|
description
optional |
String |
A description for the new field.
Example: "This field is a transformation" |
|
descriptions
optional |
Array |
A description for every of the new fields generated.
Example: ["Price 3 days ago", "Price 2 days ago", "Price 1 day ago"] |
| field | Flatline expression |
Either a json-like or lisp-like Flatline expression to generate a new field.
Example: "(* (field 5) 100)" |
| fields | Flatline expression |
Either a json-like or lisp-like Flatline expression to generate a number of fields.
Example: "(window Price -2 0)" |
|
item_analyses
optional |
Array | List of item_analyses for each of the new fields generated. |
|
item_analysis
optional |
Object |
Set of parameters to activate item analysis for the dataset. See the Section on Item Analysis for more details.
Example: |
|
label
optional |
Array |
Label of the new field.
Example: "New price" |
| labels | Array |
Labels for each of the new fields generated.
Example: ["Price-3", "Price-2", "Price-1"] |
|
name
optional |
String |
Name of the new field.
Example: "Price" |
|
names
optional |
Array |
Names for each of the new fields generated.
Example: ["P3", "P2", "P1"] |
|
optype
optional |
String |
Optype of the new field. Available optypes are "numeric", "categorical", "text", "datetime", and "items".
Example: "text" |
|
optypes
optional |
Array |
Optypes for each of the new fields generated.
Example: ["numeric", "categorical", "text"] |
|
refresh_field_type
optional |
Boolean, default is false |
Specifies whether the new field type needs to recomputed or not.
Example: true |
|
term_analyses
optional |
Array | List of term_analyses for each of the new fields generated. |
|
term_analysis
optional |
Object |
Set of parameters to activate text analysis for the dataset. See the Section on Term Analysis for more details.
Example: |
Discretization of a Continuous Field
Here's an example discretizing the "temp" field into three homogeneous levels:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb64713c192015e3000010",
"new_fields": [{
"field": "(cond (< (f \"temp\") 0) \"SOLID\"
(< (f \"temp\") 100) \"LIQUID\"
\"GAS\")",
"name":"Discrete Temp"}]}'
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b9359a3c19205ff100002a",
"new_fields": [{
"field": "(cond (> (percentile \"age\" 0.1) (f \"age\")) \"baby\"
(> (percentile \"age\" 0.2) (f \"age\")) \"child\"
(> (percentile \"age\" 0.6) (f \"age\")) \"adult\"
(> (percentile \"age\" 0.9) (f \"age\")) \"old\"
\"elder\")",
"name":"Discrete Age"}]}'Outlier Elimination
You can use, for instance, the following predicate in a filter to remove outliers:
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b9359a3c19205ff100002a",
"lisp_filter": "(< (percentile \"age\" 0.5) (f \"age\") (percentile \"age\" 0.95))"}'
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52b9359a3c19205ff100002a",
"lisp_filter": "(within-percentiles? "age" 0.5 0.95)"}'
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"origin_dataset": "dataset/52bb64713c192015e3000010",
"new_fields": [{
"field": "(if (missing? \"temp\") (mean \"temp\") (field \"temp\"))",
"name": "no missing temp"}]}'Lisp and JSON Syntaxes
Flatline also has a json-like flavor with exactly the same semantics that the lisp-like version. Basically, a Flatline formula can easily be translated to its json-like variant and vice versa by just changing parentheses to brackets, symbols to quoted strings, and adding commas to separate each sub-formula. For example, the following two formulas are the same for BigML.io.
"(/ (* 5 (- (f Fahrenheit) 32)) 9)"
["/", ["*", 5, ["-", ["f", "Fahrenheit"], 32]], 9]Final Remarks
A few important details that you should keep in mind:- Cloning a dataset implies creating also a copy of its serialized form, so you get an asyncronous resource with a status that evolves from the Summarized (4) to the Finished (5) state.
- If you specify both sampling and filtering arguments, the former are applied first.
- As with filters applied to datasources, dataset filters can use the full Flatline language to specify the boolean formula to use when sifting the input.
- Flatline performs type inference, and will in general figure out the proper optype for the generated fields, which are subsequently summarized by the dataset creation process, reaching then their final datatype (just as with a regular dataset created from a datasource). In case you need to fine-tune Flatline's inferences, you can provide an optype (or optypes) key and value in the corresponding output field entry (together with generator and names), but in general this shouldn't be needed.
- Please check the Flatline reference manual for a full description of the language for field generation and the many pre-built functions it provides.
Samples
Last Updated: Thursday, 2018-10-18 13:19
A sample provides fast-access to the raw data of a dataset on an on-demand basis.
When a new sample is requested, a copy of the dataset is stored in a special format in an in-memory cache. Multiple and different samples of the data can then be extracted using HTTPS parameterized requests by sampling sizes and simple query string filters.
Samples are ephemeral. That is to say, a sample will be available as long as GETs are requested within periods smaller than a pre-established TTL (Time to Live). The expiration timer of a sample is reset every time a new GET is received.
If requested, a sample can also perform linear regression and compute Pearson's and Spearman's correlations for either one numeric field against all other numeric fields or between two specific numeric fields.
BigML.io allows you to create, retrieve, update, delete your sample. You can also list all of your samples.
Jump to:
- Sample Base URL
- Creating a Sample
- Sample Arguments
- Retrieving a Sample
- Sample Properties
- Filtering and Paginating Fields from a Sample
- Filtering Rows from a Sample
- Updating a Sample
- Deleting a Sample
- Listing Samples
Sample Base URL
You can use the following base URL to create, retrieve, update, and delete samples. https://bigml.io/sampleAll requests to manage your samples must use HTTPS and be authenticated using your username and API key to verify your identity. See this section for more details.
Creating a Sample
To create a new sample, you need to POST to the sample base URL an object containing at least the dataset/id that you want to use to create the sample. The content-type must always be "application/json".
You can easily create a new sample using curl as follows. All you need is a valid dataset/id and your authentication variable set up as shown above.
curl "https://bigml.io/sample?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/5484b109f0a5ea59a6000018"}' BigML.io will return the newly created sample if the request succeeded.
{
"category":0,
"code":201,
"created":"2015-02-03T08:53:08.782775",
"credits":0,
"dataset":"dataset/5484b109f0a5ea59a6000018",
"description":"",
"fields_meta":{
"count":0,
"limit":1000,
"offset":0,
"total":0
},
"input_fields":[
"000000",
"000001",
"000002",
"000003",
"000004",
"000005",
"000006",
"000007",
"000008",
"000009",
"00000a",
"00000b",
"00000c",
"00000d"
],
"max_columns":14,
"max_rows":32561,
"name":"census' dataset sample",
"private":true,
"project":null,
"resource":"sample/54d9c6f4f0a5ea0b1600003a",
"seed":"c30d76cd14e24ef7ab7d28f98b3c8488",
"size":3292068,
"status":{
"code":1,
"message":"The sample is being processed and will be created soon"
},
"subscription":false,
"tags":[],
"updated":"2015-02-03T08:53:08.782792"
}
Sample Arguments
See below the full list of arguments that you can POST to create a sample.
| Argument | Type | Description |
|---|---|---|
|
category
optional |
Integer, default is the category of the dataset |
The category that best describes the sample. See the category codes for the complete list of categories.
Example: 1 |
| dataset | String |
A valid dataset/id.
Example: dataset/4f665b8103ce8920bb000006 |
|
description
optional |
String |
A description of the sample up to 8192 characters long.
Example: "This is a description of my new sample" |
|
name
optional |
String, default is dataset's name sample |
The name you want to give to the new sample.
Example: "my new sample" |
|
project
optional |
String |
The project/id you want the sample to belong to.
Example: "project/54d98718f0a5ea0b16000000" |
|
tags
optional |
Array of Strings |
A list of strings that help classify and index your sample.
Example: ["best customers", "2018"] |
|
webhook
optional |
Object |
A webhook url and an optional secret phrase. See the Section on Webhooks for more details.
Example: |
You can also use curl to customize a new sample with a name. For example, to create a new sample named "my sample" with some tags:
curl "https://bigml.io/sample?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/5484b109f0a5ea59a6000018",
"name": "my sample",
"tags": ["potential customers", "2015"]}'If you do not specify a name, BigML.io will assign to the new sample the dataset's name.
Retrieving a Sample
Each sample has a unique identifier in the form "sample/id" where id is a string of 24 alpha-numeric characters that you can use to retrieve the sample.
To retrieve a sample with curl:
curl "https://bigml.io/sample/54d9c6f4f0a5ea0b1600003a?$BIGML_AUTH"Sample Properties
Once a sample has been successfully created it will have the following properties.
| Property | Type | Description |
|---|---|---|
|
category
filterable, sortable, updatable |
Integer | One of the categories in the table of categories that help classify this resource according to the domain of application. |
| code | Integer | HTTP status code. This will be 201 upon successful creation of the sample and 200 afterwards. Make sure that you check the code that comes with the status attribute to make sure that the sample creation has been completed without errors. |
|
columns
filterable, sortable |
Integer | The number of fields returned in the sample's fields. |
|
created
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the sample was created with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
|
credits
filterable, sortable |
Float | The number of credits it cost you to create this sample. |
|
dataset
filterable, sortable |
String | The dataset/id that was used to create the sample. |
|
description
updatable |
String | A text describing the sample. It can contain restricted markdown to decorate the text. |
| fields_meta | Object | A dictionary with meta information about the fields dictionary. It specifies the total number of fields, the current offset, and limit, and the number of fields (count) returned. |
| input_fields | Array | The list of input fields' ids available to filter the sample |
| locale | String | The source's locale. |
|
max_columns
filterable, sortable |
Integer | The total number of fields in the dataset used to build the sample. |
|
max_rows
filterable, sortable |
Integer | The max number of rows in the sample. |
|
name
filterable, sortable, updatable |
String | The name of the sample as provided or based on the name of the dataset by default. |
|
private
filterable, sortable |
Boolean | Whether the sample is public or not. |
|
project
filterable, sortable, updatable |
String | The project/id the resource belongs to. |
| resource | String | The sample/id. |
|
rows
filterable, sortable |
Integer | The total number of rows in the sample, |
| sample | Object | All the information that you need to analyze the sample on your own. It includes the fields' dictionary describing the fields and their summaries and the rows. |
|
seed
filterable, sortable |
String | The string that was used to generate the sample. |
|
size
filterable, sortable |
Integer | The number of bytes of the dataset that were used to create this sample. |
| status | Object | A description of the status of the sample. It includes a code, a message, and some extra information. See the table below. |
|
subscription
filterable, sortable |
Boolean | Whether the sample was created using a subscription plan or not. |
|
tags
filterable, updatable |
Array of Strings | A list of user tags that can help classify and index this resource. |
|
updated
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the sample was updated with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
| webhook | Object | A webhook url and an optional secret phrase. See the Section on Webhooks for more details. |
A Sample Object has the following properties:
| Property | Type | Description |
|---|---|---|
|
fields
updatable |
Array | A list with an element per field in the dataset used to build the sample. Fields are paginated according to the field_meta attribute. Each entry includes the column number in the original dataset, the name of the field, the type of the field, and the summary. See this Section for more details. |
| rows | Array of Arrays | A list of lists representing the rows of the sample. Values in each list are ordered according to the fields list. |
Sample Status
Through the status field in the sample you can determine when the sample has been fully processed and ready to be used. These are the fields that a sample's status has:
| Property | Type | Description |
|---|---|---|
| code | Integer | A status code that reflects the status of the sample creation. It can be any of those that are explained here. |
| elapsed | Integer | Number of milliseconds that BigML.io took to process the sample. |
| message | String | A human readable message explaining the status. |
Once a sample has been successfully created, it will look like:
{
"category":0,
"code":200,
"columns":2,
"created":"2015-02-03T18:21:07.001000",
"credits":0,
"dataset":"dataset/5484b109f0a5ea59a6000018",
"description":"",
"fields_meta":{
"count":2,
"limit":2,
"offset":0,
"query_total":14,
"total":14
},
"input_fields":[
"000000",
"000001",
"000002",
"000003",
"000004",
"000005",
"000006",
"000007",
"000008",
"000009",
"00000a",
"00000b",
"00000c",
"00000d"
],
"locale":"en-US",
"max_columns":14,
"max_rows":32561,
"name":"my dataset",
"private":true,
"project":null,
"resource":"sample/54d9c6f4f0a5ea0b1600003a",
"rows":2,
"sample":{
"fields":[
{
"column_number":0,
"datatype":"int8",
"id":"000000",
"input_column":0,
"name":"age",
"optype":"numeric",
"order":0,
"preferred":true,
"summary":{
"bins":[
[
18.75643,
2410
],
[
21.51515,
1485
],
[
23.47642,
1675
],
[
25.48278,
1626
],
[
27.5094,
1702
],
[
29.51434,
1674
],
[
31.48252,
1716
],
[
33.50312,
1761
],
[
35.5062,
1774
],
[
37.4908,
1685
],
[
39.49317,
1610
],
[
41.49118,
1588
],
[
43.48461,
1494
],
[
46.38942,
2722
],
[
50.4325,
2252
],
[
53.47213,
879
],
[
55.46624,
785
],
[
57.50552,
724
],
[
59.46777,
667
],
[
61.46237,
558
],
[
63.47489,
438
],
[
65.45732,
328
],
[
67.4428,
271
],
[
69.45178,
197
],
[
71.48201,
139
],
[
73.44348,
115
],
[
75.50549,
91
],
[
77.44231,
52
],
[
80.28947,
76
],
[
83.95,
20
],
[
87.75,
4
],
[
90,
43
]
],
"maximum":90,
"mean":38.58165,
"median":37.03324,
"minimum":17,
"missing_count":0,
"population":32561,
"splits":[
18.58199,
20.00208,
21.38779,
22.6937,
23.89609,
25.137,
26.40151,
27.62339,
28.8206,
30.03925,
31.20051,
32.40167,
33.57212,
34.72468,
35.87617,
37.03324,
38.24651,
39.49294,
40.76573,
42.0444,
43.3639,
44.75256,
46.13703,
47.60107,
49.39145,
51.09725,
53.14627,
55.56526,
58.35547,
61.50785,
66.43583
],
"standard_deviation":13.64043,
"sum":1256257,
"sum_squares":54526623,
"variance":186.0614
}
},
{
"column_number":1,
"datatype":"string",
"id":"000001",
"input_column":1,
"name":"workclass",
"optype":"categorical",
"order":1,
"preferred":true,
"summary":{
"categories":[
[
"Private",
22696
],
[
"Self-emp-not-inc",
2541
],
[
"Local-gov",
2093
],
[
"State-gov",
1298
],
[
"Self-emp-inc",
1116
],
[
"Federal-gov",
960
],
[
"Without-pay",
14
],
[
"Never-worked",
7
]
],
"missing_count":1836
},
"term_analysis":{
"enabled":true
}
}
],
"rows":[
[
48,
"Private",
"HS-grad",
9,
"Divorced",
"Transport-moving",
"Not-in-family",
"White",
"Male",
0,
0,
65,
"United-States",
"<=50K"
],
[
71,
"Private",
"9th",
5,
"Married-civ-spouse",
"Other-service",
"Husband",
"White",
"Male",
0,
0,
40,
"United-States",
"<=50K"
]
]
},
"seed":"0493a6f8ca7aeb2aaccca22560e4b8cb",
"size":3292068,
"status":{
"code":5,
"elapsed":1,
"message":"The sample has been created",
"progress":1
},
"subscription":false,
"tags":[
"potential customers",
"2015"
],
"updated":"2015-02-03T18:21:14.537000"
}
Filtering and Paginating Fields from a Sample
A sample might be composed of hundreds or even thousands of fields. Thus when retrieving a sample, it's possible to specify that only a subset of fields be retrieved, by using any combination of the following parameters in the query string (unrecognized parameters are ignored):
Since fields is a map and therefore not ordered, the returned fields contain an additional key, order, whose integer (increasing) value gives you their ordering. In all other respects, the source is the same as the one you would get without any filtering parameter above.
The fields_meta field can help you paginate fields. Its structure is as follows:Filtering Rows from a Sample
A sample might be composed of thousands or even millions of rows. Thus when retrieving a sample, it's possible to specify that only a subset of rows be retrieved, by using any combination of the following parameters in the query string (unrecognized parameters are ignored). BigML will never return more than 1000 rows in the same response. However, you can send additional request to get different random samples.
| Parameter | Type | Description |
|---|---|---|
|
!field=
optional |
Blank |
With field the identifier of a field, select only those rows where field is not missing (i.e., it has a definite value).
Example:
|
|
!field=from,to
optional |
List |
With field the identifier of a numeric field, returns the values not in the specified interval. As with inclusion, it's possible to include or exclude the boundaries of the specified interval using square or round brackets
Example:
|
|
!field=value
optional |
List |
With field the identifier of a numeric field, returns rows for which the field doesn't equal that value.
Example:
|
|
!field=value1&!field=value2&...
optional |
String |
With field the identifier of a categorical field, select only those rows with the value of that field not one of the provided categories (when the parameter is repeated).
Example:
|
|
field=
optional |
Blank |
With field the identifier of a field, select only those rows where field is missing.
Example:
|
|
field=from,to
optional |
List |
With field the identifier of a numeric field and from, to optional numbers, specifies a filter for the numeric values of that field in the range [from, to]. One of the limits can be omitted.
Example:
|
|
field=value
optional |
List |
With field the identifier of a numeric field, returns rows for which the field equals that value.
Example:
|
|
field=value1&field=value2&...
optional |
String |
With field the identifier of a categorical field, select only those rows with the value of that field one of the provided categories (when the parameter is repeated).
Example:
|
|
index
optional |
Boolean |
When set to true, every returned row will have a first extra value which is the absolute row number, i.e., a unique row identifier. This can be useful, for instance, when you're performing various GET requests and want to compute the union of the returned regions.
Example: index=true |
|
mode
optional |
String |
One amongst deterministic, random, or linear. The way we sample the resulting rows, if needed; random means a random sample, deterministic is also random but using a fixed seed so that it's repeatable, and linear means that BigML just returns the first size rows after filtering; defaults to "deterministic".
Example: mode=random |
|
occurrence
optional |
Boolean |
When set to true, rows have prepended a value which denotes the number of times the row was present in the sample. You'll want this only when unique is set to true, otherwise all those extra values will be equal to 1. When index is also set to true (see above), the multiplicity column is added after the row index.
Example: occurrence=true |
|
precision
optional |
Integer |
The number of significant decimal numbers to keep in the returned values, for fields of type float or double. For instance, if you set precision=0, all returned numeric values will be truncated to their integral part.
Example: precision=2 |
|
row_fields
optional |
List |
You can provide a list of identifiers to be present in the samples rows, specifying which ones you actually want to see and in which order.
Example: row_fields=000000,000002 |
|
row_offset
optional |
Integer |
Skip the given number of rows. Useful when paginating over the sample in linear mode.
Example: row_offset=300 |
|
row_order_by
optional |
String |
A field that causes the returned columns to be sorted by the value of the given field, in ascending order or, when the - prefix is used, in descending order.
Example: row_order_by=-000000 |
|
rows
optional |
Integer |
The total number of rows to be returned; if less than the resulting from the rest of the filter parameters, the latter will be sampled according to mode.
Example: rows=300 |
|
seed
optional |
String |
When mode is random, you can specify your own seed in this parameter; otherwise, we choose it at random, and return the value we've used in the body of the response: that way you can make a random sampling deterministic if you happen to like a particular result.
Example: seed=mysample |
|
stat_field
optional |
String |
A field_id that corresponds to the identifier of a numeric field will cause the answer to include the Pearson's and Spearman's correlations, and linear regression terms of this field with all other numeric fields in the sample. Those values will be returned in maps keyed by "other" field id and named spearman_correlations, pearson_correlations, slopes, and intercepts.
Example: stat_field=000000 |
|
stat_fields
optional |
String |
Two field_ids that correspond to the identifiers of numeric fields will cause the answer to include the Pearson's and Spearman's correlations, and linear regression terms between the two fields. Those values will be returned in maps keyed named spearman_correlation, pearson_correlation, slope, and intercept.
Example: stat_fields=000000,000003 |
|
unique
optional |
Boolean |
When set to true, repeated rows will be removed from the sample.
Example: unique=true |
Updating a Sample
To update a sample, you need to PUT an object containing the fields that you want to update to the sample' s base URL. The content-type must always be: "application/json". If the request succeeds, BigML.io will return with an HTTP 202 response with the updated sample.
For example, to update a sample with a new name you can use curl like this:
curl "https://bigml.io/sample/54d9c6f4f0a5ea0b1600003a?$BIGML_AUTH" \
-X PUT \
-H 'content-type: application/json' \
-d '{"name": "a new name"}' Deleting a Sample
To delete a sample, you need to issue a HTTP DELETE request to the sample/id to be deleted.
Using curl you can do something like this to delete a sample:
curl -X DELETE "https://bigml.io/sample/54d9c6f4f0a5ea0b1600003a?$BIGML_AUTH"If the request succeeds you will not see anything on the command line unless you executed the command in verbose mode. Successful DELETEs will return "204 no content" responses with no body.
Once you delete a sample, it is permanently deleted. That is, a delete request cannot be undone. If you try to delete a sample a second time, or a sample that does not exist, you will receive a "404 not found" response.
However, if you try to delete a sample that is being used at the moment, then BigML.io will not accept the request and will respond with a "400 bad request" response.
See this section for more details.Listing Samples
To list all the samples, you can use the sample base URL. By default, only the 20 most recent samples will be returned. You can see below how to change this number using the limit parameter.
You can get your list of samples directly in your browser using your own username and API key with the following links.
https://bigml.io/sample?$BIGML_AUTHModels
Last Updated: Thursday, 2018-10-18 13:19
A model is a tree-like representation of your dataset with predictive power. You can create a model selecting which fields from your dataset you want to use as input fields (or predictors) and which field you want to predict, the objective field.
Each node in the model corresponds to one of the input fields. Each node has an incoming branch except the top node also known as root that has none. Each node has a number of outgoing branches except those at the bottom (the "leaves") that have none.
Each branch represents a possible value for the input field where it originates. A leaf represents the value of the objective field given all the values for each input field in the chain of branches that goes from the root to that leaf.
When you create a new model, BigML.io will automatically compute a classification model or regression model depending on whether the objective field that you want to predict is categorical or numeric, respectively.
BigML.io allows you to create, retrieve, update, delete your model. You can also list all of your models.
Jump to:
- Model Base URL
- Creating a Model
- Model Arguments
- Shuffling the Rows of Your Dataset
- Sampling Your Dataset
- Random Decision Forests
- Retrieving a Model
- Model Properties
- Filtering a Model
- PMML
- Filtering and Paginating Fields from a Model
- Updating a Model
- Deleting a Model
- Listing Models
- Weights
- Weight Field
- Objective Weights
- Automatic Balancing
Model Base URL
You can use the following base URL to create, retrieve, update, and delete models. https://bigml.io/modelAll requests to manage your models must use HTTPS and be authenticated using your username and API key to verify your identity. See this section for more details.
Creating a Model
To create a new model, you need to POST to the model base URL an object containing at least the dataset/id that you want to use to create the model. The content-type must always be "application/json".
POST /model?username=alfred;api_key=79138a622755a2383660347f895444b1eb927730; HTTP/1.1
Host: bigml.io
Content-Type: application/json
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/4f66a80803ce8940c5000006"}' BigML.io will return the newly created model if the request succeeded.
{
"category": 0,
"code": 201,
"columns": 1,
"created": "2012-11-15T02:32:48.763534",
"credits": 0.017578125,
"credits_per_prediction": 0.0,
"dataset": "dataset/50a453753c1920186d000045",
"dataset_status": true,
"description": "",
"excluded_fields": [],
"fields_meta": {
"count": 0,
"limit": 200,
"offset": 0,
"total": 0
},
"input_fields": [],
"locale": "en-US",
"max_columns": 5,
"max_rows": 150,
"missing_splits": false,
"name": "iris' dataset model",
"number_of_evaluations": 0,
"number_of_predictions": 0,
"number_of_public_predictions": 0,
"objective_field": null,
"objective_fields": [],
"ordering": 0,
"out_of_bag": false,
"price": 0.0,
"private": true,
"project": null,
"randomize": false,
"range": [
1,
150
],
"replacement": false,
"resource": "model/50a454503c1920186d000049",
"rows": 150,
"sample_rate": 1.0,
"size": 4608,
"source": "source/50a4527b3c1920186d000041",
"source_status": true,
"status": {
"code": 1,
"message": "The model is being processed and will be created soon"
},
"tags": [
"species"
],
"updated": "2012-11-15T02:32:48.763566",
"views": 0,
"white_box": false
}
Model Arguments
In addition to the dataset, you can also POST the following arguments.
| Argument | Type | Description |
|---|---|---|
|
category
optional |
Integer, default is the category of the dataset |
The category that best describes the model. See the category codes for the complete list of categories.
Example: 1 |
| dataset | String |
A valid dataset/id.
Example: dataset/4f66a80803ce8940c5000006 |
|
deep
optional |
Boolean, default is false |
Clone the dataset used to build the model while cloning the model if the dataset is available. Must be used along with the origin or shared_hash option.
Example: true |
|
depth_threshold
optional |
Integer, default is 512 |
When the depth in the tree exceeds this value, the tree stops growing. It has no effect if it's bigger than the node_threshold.
Example: 128 |
|
description
optional |
String |
A description of the model up to 8192 characters long.
Example: "This is a description of my new model" |
|
excluded_fields
optional |
Array, default is [], an empty list. None of the fields in the dataset is excluded. |
Specifies the fields that won't be included in the model.
Example: |
|
fields
optional |
Object, default is {}, an empty dictionary. That is, no names or preferred statuses are changed. |
This can be used to change the names of the fields in the model with respect to the original names in the dataset or to tell BigML that certain fields should be preferred. An entry keyed with the field id generated in the source for each field that you want the name updated.
Example: |
|
input_fields
optional |
Array, default is []. All the fields in the dataset |
Specifies the fields to be considered to create the model.
Example: |
|
max_training_time
optional |
Integer, default is 1800 |
The maximum training time allowed for the optimization, in seconds, as a strictly positive integer. Applicable only when optimize is set to true.
Example: 3600 |
|
missing_splits
optional |
Boolean, default is false |
Defines whether to explicitly include missing field values when choosing a split. When this option is enabled, generates predicates whose operators include an asterisk, such as >*, <=*, =*, or !=*. The presence of an asterisk means "or missing". So a split with the operator >* and the value 8 can be read as "x > 8 or x is missing". When using missing_splits there may also be predicates with operators = or !=, but with a null value. This means "x is missing" and "x is not missing" respectively.
Example: true |
|
name
optional |
String, default is dataset's name |
The name you want to give to the new model.
Example: "my new model" |
|
node_threshold
optional |
Integer, default is 512 |
When the number of nodes in the tree exceeds this value, the tree stops growing.
Example: 1000 |
|
number_of_model_candidates
optional |
Integer, default is 128 |
The number of model candidates evaluated over the course of the optimization. Applicable only when optimize is set to true. Maximum 200 candidates.
Example: 100 |
|
objective_field
optional |
String, default is dataset's pre-defined objective field |
Specifies the id of the field that you want to predict.
Example: "000003" |
|
objective_fields
optional |
Array, default is an array with the id of the last field in the dataset |
Specifies the id of the field that you want to predict. Even if this an array BigML.io only accepts one objective field in the current version. If both objective_field and objective_fields are specified then, objective_field takes preference.
Example: ["000003"] |
|
optimize
optional |
Boolean, default is false |
Whether the model should be built with the automatic optimization. When it is set to true, only the following modeling properties are applied: input_fields, excluded_fields, default_numeric_value, missing_numerics, sample_rate, max_training_time, objective_weights, weight_field, and number_of_model_candidates
Example: true |
|
ordering
optional |
Integer, default is 0 (deterministic). |
Specifies the type of ordering followed to build the model. There are three different types that you can specify:
Example: 1 |
| origin | String |
The model/id of the gallery model to be cloned. The price of the model must be 0 to be cloned via API. Set deep to true to clone the dataset used to build the model too.
Example: "model/5b9ab8474e172785e3000003" |
|
out_of_bag
optional |
Boolean, default is false |
Setting this parameter to true will return a sequence of the out-of-bag instances instead of the sampled instances. See the Section on Sampling for more details.
Example: true |
|
project
optional |
String |
The project/id you want the model to belong to.
Example: "project/54d98718f0a5ea0b16000000" |
|
random_candidate_ratio
optional |
Float |
A real number between 0 and 1. When randomize is true and random_candidate_ratio is given, BigML randomizes the tree and uses random_candidate_ratio * total fields (counting the number of terms in text fields as fields). To get the final number of candidate fields we round down to the nearest integer, but if the result is 0 we'll use 1 instead. If both random_candidates and random_candidate_ratio are given, BigML ignores random_candidate_ratio.
Example: 0.2 |
|
random_candidates
optional |
Integer, default is the square root of the total number of input fields. |
Sets the number of random fields considered when randomize is true.
Example: 10 |
|
randomize
optional |
Boolean, default is false |
Setting this parameter to true will consider only a subset of the possible fields when choosing a split. See the Section on Random Decision Forests below.
Example: true |
|
range
optional |
Array, default is [1, max rows in the dataset] |
The range of successive instances to build the model.
Example: [1, 150] |
|
replacement
optional |
Boolean, default is false |
Whether sampling should be performed with or without replacement. See the Section on Sampling for more details.
Example: true |
|
sample_rate
optional |
Float, default is 1.0 |
A real number between 0 and 1 specifying the sample rate. See the Section on Sampling for more details.
Example: 0.5 |
|
seed
optional |
String |
A string to be hashed to generate deterministic samples. See the Section on Sampling for more details.
Example: "MySample" |
| shared_hash | String |
The shared hash of the shared model to be cloned. The price of the model must be 0 to be cloned via API.
Example: "kpY46mNuNVReITw0Z1mAqoQ9ySW" |
|
split_candidates
optional |
Integer, default is 32 |
The number of split points that are considered whenever the tree evaluates a numeric field. Minimum 1 and maximum 1024
Example: 128 |
|
stat_pruning
optional |
Boolean |
Activates statistical pruning on your decision tree model.
Example: true |
|
support_threshold
optional |
Float, default is 0 |
The parameter controls the minimum amount of support each child node must contain to be valid as a possible split. So, if it is 3, then a both children of a new split must have 3 instances supporting them. Since instances may have non-integer weights, non-integer values are valid.
Example: 16 |
|
tags
optional |
Array of Strings |
A list of strings that help classify and index your model.
Example: ["best customers", "2018"] |
|
webhook
optional |
Object |
A webhook url and an optional secret phrase. See the Section on Webhooks for more details.
Example: |
You can also use curl to customize a new model. For example, to create a new model named "my model", with only certain rows, and with only three fields:
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/4f66a80803ce8940c5000006",
"input_fields": ["000001", "000003"],
"name": "my model",
"range": [25, 125]}'If you do not specify a name, BigML.io will assign to the new model the dataset's name. If you do not specify a range of instances, BigML.io will use all the instances in the dataset. If you do not specify any input fields, BigML.io will include all the input fields in the dataset, and if you do not specify an objective field, BigML.io will use the last field in your dataset.
Shuffling the Rows of Your Dataset
By default, rows from the input dataset are deterministically shuffled before being processed, to avoid inaccurate models caused by ordered fields in the input rows. Since the shuffling is deterministic, i.e., always the same for a given dataset, retraining a model for the same dataset will always yield the same result.
However, you can modify this default behaviour by including the ordering argument in the model creation request, where "ordering" here is a shortcut for "ordering for the traversal of input rows". When this property is absent or set to 0, deterministic shuffling takes place; otherwise, you can set it to:
- Linear: If you know that your input is already in random order. Setting "ordering" to 1 in your model request tells BigML to traverse the dataset in a linear fashion, without performing any shuffling (and therefore operating faster).
- Random: If you'd like to perform a really random shuffling, most probably different from any other one attempted before. Setting "ordering" to 2 will shuffle the input rows non-deterministically.
Sampling Your Dataset
You can limit the dataset rows that are used to create a model in two ways (which can be combined), namely, by specifying a row range and by asking for a sample of the (alreaday clipped) input rows.
The row range is specified with the range argument defined in the Section on Arguments above.
To specify a sample, which is taken over the row range or over the whole dataset if a range is not provided, you can add the following arguments to the creation request:
- sample_rate : A positive number that specifies the sampling rate, i.e., how often we pick a row from the range. In other words, the final number of rows will be the size of the range multiply by the sample_rate, unless "out_of_bag" is true (see below).
- replacement : A boolean indicating whether sampling should be performed with or without replacement, i.e., the same instance may be selected multiple times for inclusion in the result set. Defaults to false.
- out_of_bag : If an instance isn't selected as part of a sampling, it's called out of bag. Setting this parameter to true will return a sequence of the out-of-bag instances instead of the sampled instances. This can be useful when paired with "seed". When replacement is false, the final number of row returned is the size of the range multiply by one minus the sample_rate. Out-of-bag sampling with replacement gives rise to variable-size samples. Defaults to false.
- seed : Rows are sampled probabilistically using a random string, which means that, in general, two identical samples of the same row range of the same dataset will be different. If you provide a seed (as an arbitrary string), its hash value will be used as the seed, and it'll be possible for you to generate deterministic samples.
Finally, note that the "ordering" of the dataset described in the previous subsection is used on the result of the sampling.
Here's an example of a model request with range and sampling specifications:
curl https://bigml.io/model?$BIGML_AUTH \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/505f43223c1920eccc000297", "range": [1, 5000], "sample_rate": 0.5, "replacement": true}'
Random Decision Forests
A model can be randomized by setting the randomize parameter to true. The default is false.
When randomized, the model considers only a subset of the possible fields when choosing a split. The size of the subset will be the square root of the total number of input fields. So if there are 100 input fields, each split will only consider 10 fields randomly chosen from the 100. Every split will choose a new subset of fields.
Although randomize could be used for other purposes, it's intended for growing random decision forests. To grow tree models for a random forest, set randomize to true and select a sample from the dataset. Traditionally this is a 1.0 sample rate with replacement, but we suggest a 0.63 sample rate without replacement.
Retrieving a Model
Each model has a unique identifier in the form "model/id" where id is a string of 24 alpha-numeric characters that you can use to retrieve the model.
To retrieve a model with curl:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH"You can also use your browser to visualize the model using the full BigML.io URL or pasting the model/id into the BigML.com dashboard.
Model Properties
Once a model has been successfully created it will have the following properties.
| Property | Type | Description |
|---|---|---|
|
boosted_ensemble
filterable, sortable |
Boolean | Whether the model was built as part of an ensemble with boosted trees. |
| boosting | Object |
Boosting attribute for the boosted tree. See the Gradient Boosting section for more information.
Example: |
|
category
filterable, sortable, updatable |
Integer | One of the categories in the table of categories that help classify this resource according to the domain of application. |
| code | Integer | HTTP status code. This will be 201 upon successful creation of the model and 200 afterwards. Make sure that you check the code that comes with the status attribute to make sure that the model creation has been completed without errors. |
|
columns
filterable, sortable |
Integer | The number of fields in the model. |
|
composites
filterable, sortable |
Array of Strings | The list of composite ids that reference this model. |
|
created
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the model was created with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
|
credits
filterable, sortable |
Float | The number of credits it cost you to create this model. |
|
credits_per_prediction
filterable, sortable, updatable |
Float | This is the number of credits that other users will consume to make a prediction with your model if you made it public. |
|
dataset
filterable, sortable |
String | The dataset/id that was used to build the model. |
| dataset_field_types | Object | A dictionary that informs about the number of fields of each type in the dataset used to create the model. It has an entry per each field type (categorical, datetime, numeric, and text), an entry for preferred fields, and an entry for the total number of fields. |
|
dataset_status
filterable, sortable |
Boolean | Whether the dataset is still available or has been deleted. |
| deep | Boolean | Whether the dataset used to build the original model is also requested to be cloned or not. |
|
description
updatable |
String | A text describing the model. It can contain restricted markdown to decorate the text. |
|
ensemble
filterable, sortable |
Boolean | Whether the model was built as part of an ensemble of not. |
|
ensemble_id
filterable, sortable |
String | The ensemble id. |
|
ensemble_index
filterable, sortable |
Integer | The number of order in the ensemble. |
| excluded_fields | Array | The list of fields's ids that were excluded to build the model. |
| fields_meta | Object | A dictionary with meta information about the fields dictionary. It specifies the total number of fields, the current offset, and limit, and the number of fields (count) returned. |
|
fusions
filterable, sortable |
Array of Strings | The list of fusion ids that reference this model. |
| input_fields | Array | The list of input fields' ids used to build the models of the model. |
| locale | String | The dataset's locale. |
|
max_columns
filterable, sortable |
Integer | The total number of fields in the dataset used to build the model. |
|
max_rows
filterable, sortable |
Integer | The maximum number of instances in the dataset that can be used to build the model. |
| max_training_time | Integer | The maximum training time allowed for the optimization, in seconds. |
|
missing_splits
filterable, sortable |
Boolean | Whether to explicitly include missing field values when choosing a split while growing a model. |
| model | Object | All the information that you need to recreate or use the model on your own. It includes a very intuitive description of the tree-like structure that makes the model up and the field's dictionary describing the fields and their summaries. |
|
name
filterable, sortable, updatable |
String | The name of the model as your provided or based on the name of the dataset by default. |
|
node_threshold
filterable, sortable |
String | The maximum number of nodes that the model will grow. |
|
number_of_batchpredictions
filterable, sortable |
Integer | The current number of batch predictions that use this model. |
|
number_of_evaluations
filterable, sortable |
Integer | The current number of evaluations that use this model. |
| number_of_model_candidates | Integer | The number of model candidates evaluated over the course of the optimization. |
|
number_of_predictions
filterable, sortable |
Integer | The current number of predictions that use this model. |
|
number_of_public_predictions
filterable, sortable |
Integer | The current number of public predictions that use this model. |
| objective_field | String | The id of the field that the model predicts. |
| objective_fields | Array | Specifies the list of ids of the field that the model predicts. Even if this is an array BigML.io only accepts one objective field in the current version. |
| optimize | Boolean | Whether the model was built with the automatic optimization. |
|
optiml
filterable, sortable |
String | The optiml/id that created this model. |
|
optiml_status
filterable, sortable |
Boolean | Whether the OptiML is still available or has been deleted. |
|
ordering
filterable, sortable |
Integer |
The order used to chose instances from the dataset to build the model. There are three different types:
|
|
origin
filterable, sortable |
String | The model/id of the original gallery model. |
|
out_of_bag
filterable, sortable |
Boolean | Whether the out-of-bag instances were used to create the model instead of the sampled instances. |
|
price
filterable, sortable, updatable |
Float | The price other users must pay to clone your model. |
|
private
filterable, sortable, updatable |
Boolean | Whether the model is public or not. |
|
project
filterable, sortable, updatable |
String | The project/id the resource belongs to. |
|
random_candidate_ratio
filterable, sortable |
Float | The random candidate ratio considered when randomize is true. |
|
random_candidates
filterable, sortable |
Integer | The number of random fields considered when randomize is true. |
|
randomize
filterable, sortable |
Boolean | Whether the model splits considered only a random subset of the fields or all the fields available. |
| range | Array | The range of instances used to build the model. |
|
replacement
filterable, sortable |
Boolean | Whether the instances sampled to build the model were selected using replacement or not. |
| resource | String | The model/id. |
|
rows
filterable, sortable |
Integer | The total number of instances used to build the model |
|
sample_rate
filterable, sortable |
Float | The sample rate used to select instances from the dataset to build the model. |
|
seed
filterable, sortable |
String | The string that was used to generate the sample. |
|
selective_pruning
filterable, sortable |
Boolean | If true, selective pruning throttled the strength of the statistical pruning depending on the size of the dataset. |
|
shared
filterable, sortable, updatable |
Boolean | Whether the model is shared using a private link or not. |
|
shared_clonable
filterable, sortable, updatable |
Boolean | Whether the shared model can be cloned or not. |
| shared_hash | String | The hash that gives access to this model if it has been shared using a private link. |
| sharing_key | String | The alternative key that gives read access to this model. |
|
size
filterable, sortable |
Integer | The number of bytes of the dataset that were used to create this model. |
|
source
filterable, sortable |
String | The source/id that was used to build the dataset. |
|
source_status
filterable, sortable |
Boolean | Whether the source is still available or has been deleted. |
|
split_candidates
filterable, sortable |
Integer | The number of split points that are considered whenever the tree evaluates a numeric field. Minimum 1 and maximum 1024 |
|
stat_pruning
filterable, sortable |
Boolean | Whether statistical pruning was used when building the model. |
| status | Object | A description of the status of the model. It includes a code, a message, and some extra information. See the table below. |
|
subscription
filterable, sortable |
Boolean | Whether the model was created using a subscription plan or not. |
|
support_threshold
filterable, sortable |
Float | The parameter controls the minimum amount of support each child node must contain to be valid as a possible split. |
|
tags
filterable, updatable |
Array of Strings | A list of user tags that can help classify and index this resource. |
|
updated
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the model was updated with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
| webhook | Object | A webhook url and an optional secret phrase. See the Section on Webhooks for more details. |
|
white_box
filterable, sortable |
Boolean | Whether the model is publicly shared as a white-box. |
A Model Object has the following properties:
| Property | Type | Description |
|---|---|---|
| depth_threshold | Integer | The depth, or generation, limit for a tree. |
| distribution | Object | This dictionary gives information about how the training data is distributed across the tree leaves. More concretely, it contains the training data distribution with key training, and the distribution for the actual prediction values of the tree with key predictions. The former is just the objective_summary of the tree root (see below), copied for easier individual retrieval, and both have the format of the objective summary in the tree nodes. |
|
fields
updatable |
Object | A dictionary with an entry per field in the dataset used to build the model. Fields are paginated according to the field_meta attribute. Each entry includes the column number in original source, the name of the field, the type of the field, and the summary. See this Section for more details. |
| importance | Array of Arrays | A list of pairs [field_id, importance]. Importance is the amount by which each field in the model reduces prediction error, normalized to be between zero and one. Note that fields with an importance of zero may still be correlated with the objective; they were just not used in the model. |
| kind | String | The type of model. Currently, only stree is supported. |
| missing_strategy | String | Default strategy followed by the model when it finds a missing value. Currently, last_prediction. At prediction time you can opt for using proportional. See this Section for more details. |
| model_fields | Object | A dictionary with an entry per field used by the model (not all the fields that were available in the dataset). They follow the same structure as the fields attribute above except that the summary is not present. |
| root | Object | A Node Object, a tree-like recursive structure representing the model. |
| split_criterion | Integer | Method of choosing best attribute and split point for a given node. DEPRECATED |
| support_threshold | Float | A number between 0 and 1. For a split to be valid, each child's support (instances / total instances) must be greater than this threshold. |
Node Objects have the following properties:
| Property | Type | Description |
|---|---|---|
| children | Array | Array of Node Objects. |
| confidence | Float | For classification models, a number between 0 and 1 that expresses how certain the model is of the prediction. For regression models, a number mapped to the top end of a 95 confidence interval around the expected error at that node (measured using the variance of the output at the node). See the Section on Confidence for more details. Note that for models you might have created using the first versions of BigML this value might be null. |
| count | Integer | Number of instances classfied by this node. |
| objective_summary | Object | An Objective Summary Object summarizes the objective field's distribution at this node. |
| output | Number or String | Prediction at this node. |
| predicate | Boolean or Object | Predicate structure to make a decision at this node. |
Objective Summary Objects have the following properties:
| Property | Type | Description |
|---|---|---|
| bins | Array | If the objective field is numeric and the number of distinct values is greater than 32. An array that represents an approximate histogram of the distribution. It consists of value pairs, where the first value is the mean of a histogram bin and the second value is the bin population. For more information, see our blog post or read this paper. |
| categories | Array | If the objective field is categorical, an array of pairs where the first element of each pair is one of the unique categories and the second element is the count for that category. |
| counts | Array | If the objective field is numeric and the number of distinct values is less than or equal to 32, an array of pairs where the first element of each pair is one of the unique values found in the field and the second element is the count. |
| maximum | Number | The maximum of the objective field's values. Available when 'bins' is present. |
| minimum | Number | The minimum of the objective field's values. Available when 'bins' is present. |
Predicate Objects have the following properties:
Model Status
Creating a model is a process that can take just a few seconds or a few days depending on the size of the dataset used as input and on the workload of BigML's systems. The model goes through a number of states until its fully completed. Through the status field in the model you can determine when the model has been fully processed and ready to be used to create predictions. These are the properties that a model's status has:
| Property | Type | Description |
|---|---|---|
| code | Integer | A status code that reflects the status of the model creation. It can be any of those that are explained here. |
| elapsed | Integer | Number of milliseconds that BigML.io took to process the model. |
| message | String | A human readable message explaining the status. |
| progress | Float, between 0 and 1 | How far BigML.io has progressed building the model. |
Once a model has been successfully created, it will look like:
{
"category": 0,
"code": 200,
"columns": 5,
"created": "2012-11-15T02:32:48.763000",
"credits": 0.017578125,
"credits_per_prediction": 0.0,
"dataset": "dataset/50a453753c1920186d000045",
"dataset_status": true,
"description": "",
"excluded_fields": [],
"fields_meta": {
"count": 5,
"limit": 200,
"offset": 0,
"total": 5
},
"input_fields": [
"000000",
"000001",
"000002",
"000003"
],
"locale": "en_US",
"max_columns": 5,
"max_rows": 150,
"missing_splits": false,
"model": {
"depth_threshold": 20,
"fields": {
"000000": {
"column_number": 0,
"datatype": "double",
"name": "sepal length",
"optype": "numeric",
"order": 0,
"preferred": true,
"summary": {
"bins": [
[
4.3,
1
],
[
4.425,
4
],
[
4.6,
4
],
[
4.7,
2
],
[
4.8,
5
],
[
4.9,
6
],
[
5,
10
],
[
5.1,
9
],
[
5.2,
4
],
[
5.3,
1
],
[
5.4,
6
],
[
5.5,
7
],
[
5.6,
6
],
[
5.7,
8
],
[
5.8,
7
],
[
5.9,
3
],
[
6,
6
],
[
6.1,
6
],
[
6.2,
4
],
[
6.3,
9
],
[
6.44167,
12
],
[
6.6,
2
],
[
6.7,
8
],
[
6.8,
3
],
[
6.92,
5
],
[
7.1,
1
],
[
7.2,
3
],
[
7.3,
1
],
[
7.4,
1
],
[
7.6,
1
],
[
7.7,
4
],
[
7.9,
1
]
],
"maximum": 7.9,
"mean": 5.84333,
"median": 5.77889,
"minimum": 4.3,
"missing_count": 0,
"population": 150,
"splits": [
4.51526,
4.67252,
4.81113,
4.89582,
4.96139,
5.01131,
5.05992,
5.11148,
5.18177,
5.35681,
5.44129,
5.5108,
5.58255,
5.65532,
5.71658,
5.77889,
5.85381,
5.97078,
6.05104,
6.13074,
6.23023,
6.29578,
6.35078,
6.41459,
6.49383,
6.63013,
6.70719,
6.79218,
6.92597,
7.20423,
7.64746
],
"standard_deviation": 0.82807,
"sum": 876.5,
"sum_squares": 5223.85,
"variance": 0.68569
}
},
"000001": {
"column_number": 1,
"datatype": "double",
"name": "sepal width",
"optype": "numeric",
"order": 1,
"preferred": true,
"summary": {
"counts": [
[
2,
1
],
[
2.2,
3
],
[
2.3,
4
],
[
2.4,
3
],
[
2.5,
8
],
[
2.6,
5
],
[
2.7,
9
],
[
2.8,
14
],
[
2.9,
10
],
[
3,
26
],
[
3.1,
11
],
[
3.2,
13
],
[
3.3,
6
],
[
3.4,
12
],
[
3.5,
6
],
[
3.6,
4
],
[
3.7,
3
],
[
3.8,
6
],
[
3.9,
2
],
[
4,
1
],
[
4.1,
1
],
[
4.2,
1
],
[
4.4,
1
]
],
"maximum": 4.4,
"mean": 3.05733,
"median": 3.02044,
"minimum": 2,
"missing_count": 0,
"population": 150,
"standard_deviation": 0.43587,
"sum": 458.6,
"sum_squares": 1430.4,
"variance": 0.18998
}
},
"000002": {
"column_number": 2,
"datatype": "double",
"name": "petal length",
"optype": "numeric",
"order": 2,
"preferred": true,
"summary": {
"bins": [
[
1,
1
],
[
1.1,
1
],
[
1.2,
2
],
[
1.3,
7
],
[
1.4,
13
],
[
1.5,
13
],
[
1.63636,
11
],
[
1.9,
2
],
[
3,
1
],
[
3.3,
2
],
[
3.5,
2
],
[
3.6,
1
],
[
3.75,
2
],
[
3.9,
3
],
[
4.0375,
8
],
[
4.23333,
6
],
[
4.46667,
12
],
[
4.6,
3
],
[
4.74444,
9
],
[
4.94444,
9
],
[
5.1,
8
],
[
5.25,
4
],
[
5.46,
5
],
[
5.6,
6
],
[
5.75,
6
],
[
5.95,
4
],
[
6.1,
3
],
[
6.3,
1
],
[
6.4,
1
],
[
6.6,
1
],
[
6.7,
2
],
[
6.9,
1
]
],
"maximum": 6.9,
"mean": 3.758,
"median": 4.34142,
"minimum": 1,
"missing_count": 0,
"population": 150,
"splits": [
1.25138,
1.32426,
1.37171,
1.40962,
1.44567,
1.48173,
1.51859,
1.56301,
1.6255,
1.74645,
3.23033,
3.675,
3.94203,
4.0469,
4.18243,
4.34142,
4.45309,
4.51823,
4.61771,
4.72566,
4.83445,
4.93363,
5.03807,
5.1064,
5.20938,
5.43979,
5.5744,
5.6646,
5.81496,
6.02913,
6.38125
],
"standard_deviation": 1.7653,
"sum": 563.7,
"sum_squares": 2582.71,
"variance": 3.11628
}
},
"000003": {
"column_number": 3,
"datatype": "double",
"name": "petal width",
"optype": "numeric",
"order": 3,
"preferred": true,
"summary": {
"counts": [
[
0.1,
5
],
[
0.2,
29
],
[
0.3,
7
],
[
0.4,
7
],
[
0.5,
1
],
[
0.6,
1
],
[
1,
7
],
[
1.1,
3
],
[
1.2,
5
],
[
1.3,
13
],
[
1.4,
8
],
[
1.5,
12
],
[
1.6,
4
],
[
1.7,
2
],
[
1.8,
12
],
[
1.9,
5
],
[
2,
6
],
[
2.1,
6
],
[
2.2,
3
],
[
2.3,
8
],
[
2.4,
3
],
[
2.5,
3
]
],
"maximum": 2.5,
"mean": 1.19933,
"median": 1.32848,
"minimum": 0.1,
"missing_count": 0,
"population": 150,
"standard_deviation": 0.76224,
"sum": 179.9,
"sum_squares": 302.33,
"variance": 0.58101
}
},
"000004": {
"column_number": 4,
"datatype": "string",
"name": "species",
"optype": "categorical",
"order": 4,
"preferred": true,
"summary": {
"categories": [
[
"Iris-versicolor",
50
],
[
"Iris-setosa",
50
],
[
"Iris-virginica",
50
]
],
"missing_count": 0
}
}
},
"importance": [
[
"000002",
0.53159
],
[
"000003",
0.4633
],
[
"000000",
0.00511
],
[
"000001",
0
]
],
"kind": "stree",
"missing_strategy": "last_prediction",
"model_fields": {
"000000": {
"column_number": 0,
"datatype": "double",
"name": "sepal length",
"optype": "numeric",
"order": 0,
"preferred": true
},
"000001": {
"column_number": 1,
"datatype": "double",
"name": "sepal width",
"optype": "numeric",
"order": 1,
"preferred": true
},
"000002": {
"column_number": 2,
"datatype": "double",
"name": "petal length",
"optype": "numeric",
"order": 2,
"preferred": true
},
"000003": {
"column_number": 3,
"datatype": "double",
"name": "petal width",
"optype": "numeric",
"order": 3,
"preferred": true
},
"000004": {
"column_number": 4,
"datatype": "string",
"name": "species",
"optype": "categorical",
"order": 4,
"preferred": true
}
},
"root": {
"children": [
{
"confidence": 0.92865,
"count": 50,
"objective_summary": {
"categories": [
[
"Iris-setosa",
50
]
]
},
"output": "Iris-setosa",
"predicate": {
"field": "000002",
"operator": "<=",
"value": 2.45
}
},
{
"children": [
{
"children": [
{
"children": [
{
"children": [
{
"children": [
{
"confidence": 0.34237,
"count": 2,
"objective_summary": {
"categories": [
[
"Iris-virginica",
2
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000000",
"operator": ">",
"value": 5.95
}
},
{
"confidence": 0.20654,
"count": 1,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
1
]
]
},
"output": "Iris-versicolor",
"predicate": {
"field": "000000",
"operator": "<=",
"value": 5.95
}
}
],
"confidence": 0.20765,
"count": 3,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
1
],
[
"Iris-virginica",
2
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000000",
"operator": "<=",
"value": 6.4
}
},
{
"confidence": 0.20654,
"count": 1,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
1
]
]
},
"output": "Iris-versicolor",
"predicate": {
"field": "000000",
"operator": ">",
"value": 6.4
}
}
],
"confidence": 0.15004,
"count": 4,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
2
],
[
"Iris-virginica",
2
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000001",
"operator": ">",
"value": 2.9
}
},
{
"confidence": 0.60966,
"count": 6,
"objective_summary": {
"categories": [
[
"Iris-virginica",
6
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000001",
"operator": "<=",
"value": 2.9
}
}
],
"confidence": 0.49016,
"count": 10,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
2
],
[
"Iris-virginica",
8
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000002",
"operator": "<=",
"value": 5.05
}
},
{
"confidence": 0.90819,
"count": 38,
"objective_summary": {
"categories": [
[
"Iris-virginica",
38
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000002",
"operator": ">",
"value": 5.05
}
}
],
"confidence": 0.86024,
"count": 48,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
2
],
[
"Iris-virginica",
46
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000003",
"operator": ">",
"value": 1.65
}
},
{
"children": [
{
"confidence": 0.92444,
"count": 47,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
47
]
]
},
"output": "Iris-versicolor",
"predicate": {
"field": "000002",
"operator": "<=",
"value": 4.95
}
},
{
"children": [
{
"confidence": 0.43849,
"count": 3,
"objective_summary": {
"categories": [
[
"Iris-virginica",
3
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000000",
"operator": ">",
"value": 6.05
}
},
{
"children": [
{
"confidence": 0.20654,
"count": 1,
"objective_summary": {
"categories": [
[
"Iris-virginica",
1
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000003",
"operator": "<=",
"value": 1.55
}
},
{
"confidence": 0.20654,
"count": 1,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
1
]
]
},
"output": "Iris-versicolor",
"predicate": {
"field": "000003",
"operator": ">",
"value": 1.55
}
}
],
"confidence": 0.09453,
"count": 2,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
1
],
[
"Iris-virginica",
1
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000000",
"operator": "<=",
"value": 6.05
}
}
],
"confidence": 0.37553,
"count": 5,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
1
],
[
"Iris-virginica",
4
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000002",
"operator": ">",
"value": 4.95
}
}
],
"confidence": 0.81826,
"count": 52,
"objective_summary": {
"categories": [
[
"Iris-virginica",
4
],
[
"Iris-versicolor",
48
]
]
},
"output": "Iris-versicolor",
"predicate": {
"field": "000003",
"operator": "<=",
"value": 1.65
}
}
],
"confidence": 0.40383,
"count": 100,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
50
],
[
"Iris-virginica",
50
]
]
},
"output": "Iris-virginica",
"predicate": {
"field": "000002",
"operator": ">",
"value": 2.45
}
}
],
"confidence": 0.26289,
"count": 150,
"objective_summary": {
"categories": [
[
"Iris-versicolor",
50
],
[
"Iris-setosa",
50
],
[
"Iris-virginica",
50
]
]
},
"output": "Iris-virginica",
"predicate": true
},
"split_criterion": "information_gain_mix",
"support_threshold": 0
},
"name": "iris' dataset model",
"number_of_evaluations": 0,
"number_of_predictions": 0,
"number_of_public_predictions": 0,
"objective_field": "000004",
"objective_fields": [
"000004"
],
"ordering": 0,
"out_of_bag": false,
"price": 0.0,
"private": true,
"project": null,
"randomize": false,
"range": [
1,
150
],
"replacement": false,
"resource": "model/50a454503c1920186d000049",
"rows": 150,
"sample_rate": 1.0,
"selective_pruning": true,
"size": 4608,
"source": "source/50a4527b3c1920186d000041",
"source_status": true,
"stat_pruning": true,
"status": {
"code": 5,
"elapsed": 413,
"message": "The model has been created",
"progress": 1.0
},
"tags": [
"species"
],
"updated": "2012-11-15T02:32:50.149000",
"views": 0,
"white_box": false
}
Filtering a Model
It is possible to filter the tree returned by a GET to the model location by means of two optional query string parameters, namely support and value.
Filter by Support
Support is a number from 0 to 1 that specifies the minimum fraction of the total number of instances that a given branch must cover to be retained in the resulting tree. Thus, asking for (minimum) support of 0, is just asking for the whole tree, while something like:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;support=1.0"will return just the root node, that being the only one that covers all instances. If you repeat the support parameter in the query string, the last one is used. Non-parseable support values are ignored.
Filter by Values and Value Intervals
Value is a concrete value or interval of values (for regression trees) that a leaf must predict to be kept in the returning tree. For instance:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=Iris-setosa"will return only those branches in the tree whose leaves predict "Iris-setosa" as the value of the (categorical) objective field, while something like:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=[10,20]"for a regression model will include only those leaves predicting an objective value between 10 and 20. You can also specify sharp values for regression models:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=23.2"will retrieve only those branches whose predictions are exactly 23.2. It is possible to specify multiple values, as in:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=Iris-setosa&value=Iris-versicolor"
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=(10,20]&value=[-1.234,3.3)"
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=(10.2,20)&value=28.1&value=0.1"in which case the union of the different predicates is used (i.e., the first query will return a tree will all leaves predicting "Iris-setosa" and all leaves predicting "Iris-versicolor".
Intervals can be closed or open in either end. For example, "(-2,10]", "[1,2)" or "(-1.234,0)", and the values of the left or right limits can be omitted, in which case they're taken as negative and positive infinity, respectively; thus "(,3]" denotes all values less or equal to three, as does "[,3]" (infinity not being a valid value for a numeric prediction), while "(0,)" accepts any positive value.
Filter by Confidence / Probability / Expected Error
Confidence is a concrete value or interval of values that a leaf must have to be kept in the returning tree. The specification of intervals follows the same conventions as those of value. Since confidences are a continuous value, the most common case will be asking for a range, but the service will accept also individual values. It's also possible to specify both a value and a confidence. For instance:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;value=Iris-setosa&confidence=[0.3,]"asks for a tree with only those leaves that predict "Iris-setosa" with a confidence greater or equal to 0.3, while
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;confidence=[,0.25)"returns a model where all leaves with confidence strictly less than 0.25. Confidence filters will work both for classification regression problems, since we call the regression expected error confidence in our JSON. If desired (and only for regression), one can specify a filter using expected_error instead:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;expected_error=[,0.25)"If you specify both confidence and expected_error, only one of them will be used: confidence for classifications, expected_error for regressions. If only confidence is specified, it will always be used (confidence is an alias for the the expected error in regressions). If only expected_error is specified, it will only be used if the model is a regression.
Filters by probability works exactly as filters by confidence, but replacing probability for confidence. As a consequence, they'll only have an effect on classification problems.
Finally, note that it is also possible to specify support, value, confidence, probability, and expected_error parameters in the same query.
PMML
The default model output format is JSON. However, the pmml parameter allows to include a PMML version of the model. The model will include a XML document that fullfils PMML v4.1. For example:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;pmml=yes"
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH;pmml=only"Filtering and Paginating Fields from a Model
A model might be composed of hundreds or even thousands of fields. Thus when retrieving a model, it's possible to specify that only a subset of fields be retrieved, by using any combination of the following parameters in the query string (unrecognized parameters are ignored):
Since fields is a map and therefore not ordered, the returned fields contain an additional key, order, whose integer (increasing) value gives you their ordering. In all other respects, the source is the same as the one you would get without any filtering parameter above.
The fields_meta field can help you paginate fields. Its structure is as follows:Updating a Model
To update a model, you need to PUT an object containing the fields that you want to update to the model' s base URL. The content-type must always be: "application/json". If the request succeeds, BigML.io will return with an HTTP 202 response with the updated model.
For example, to update a model with a new name you can use curl like this:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH" \
-X PUT \
-d '{"name": "a new name"}' \
-H 'content-type: application/json'If you want to update a model with a new label and description for a specific field you can use curl like this:
curl "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH" \
-X PUT \
-d '{"fields": {"000000": {"label": "a longer name", "description": "an even longer description"}}}' \
-H 'content-type: application/json'Deleting a Model
To delete a model, you need to issue a HTTP DELETE request to the model/id to be deleted.
Using curl you can do something like this to delete a model:
curl -X DELETE "https://bigml.io/model/50a454503c1920186d000049?$BIGML_AUTH"If the request succeeds you will not see anything on the command line unless you executed the command in verbose mode. Successful DELETEs will return "204 no content" responses with no body.
Once you delete a model, it is permanently deleted. That is, a delete request cannot be undone. If you try to delete a model a second time, or a model that does not exist, you will receive a "404 not found" response.
However, if you try to delete a model that is being used at the moment, then BigML.io will not accept the request and will respond with a "400 bad request" response.
See this section for more details.Listing Models
To list all the models, you can use the model base URL. By default, only the 20 most recent models will be returned. You can see below how to change this number using the limit parameter.
You can get your list of models directly in your browser using your own username and API key with the following links.
https://bigml.io/model?$BIGML_AUTHWeights
BigML.io has added three new ways in which you can use weights to deal with imbalanced datasets:
- Weight Field: considering the values one of the fields in the dataset as weight for the instances. This is valid for both regression and classification models.
- Objective Weights: submitting a specific weight for each class in classification models.
- Automatic Balancing: setting the balance argument to true to let BigML automatically balance all the classes evenly.
Let's see each method in more detail.
Weight Field
A weight_field may be declared for either regression or classification models. Any numeric field with no negative or missing values is valid as a weight field. Each instance will be weighted individually according to the weight field's value. See the toy dataset for credit card transactions below.
online, transaction, pending transactions, days since last transaction, distance, transactions today, balance, mtd, fraud, weight
yes, 10, 3, 31, low, 3, -3250, -1500, no, 1
no, 20, 30, 1, high, 0, 0, -300, no, 1
no, 40, 13, 210, low, 1, -19890, -30, no, 1
yes, 500, 0, 1, high, 0, 0, 0, yes, 10
no, 10, 1, 32, low, 0, -2500, -7891, no, 1
yes, 100, 0, 3, low, 0, -5194, -120, no, 1
yes, 100, 1, 4, low, 0, 0, 1500, no, 1
yes, 1000, 0, 1, high, 0, 0, 0, yes, 10
no, 150, 3, 1, low, 5, -3250, 1500, no, 1
no, 75, 5, 1, high, 1, -3250, 1500, no, 1
yes, 10, 23, 0, low, 1, -3250, 1500, no, 1
yes, 10, 3, 31, low, 3, -3250, -1500, no, 1
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/52bcdc513c1920e4a300006e",
"objective_field": "000008",
"weight_field": "000009"
}'With Flatline, you can define arbitrarily complex functions to produce weight fields, making this the most flexible and powerful way to produce weighted models.
For instance, the request below would create a new dataset using the example above that will add a new weight field using the previous and multiplying by two when the amount of the transaction is higher than 500.
curl "https://bigml.io/dataset?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/52bcdc513c1920e4a300006e",
"new_fields": [{
"field": "(if (and (= (f fraud) \"yes\") (> (f transaction) 500)) (* (f weight) 2) (f weight))",
"name": "new weight"}]
}'Objective Weights
The second method for adding weights only applies to classification models. A set of objective_weights may be defined, one per objective class. Each instance will be weighted according to its class weight.
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/52bcdc513c1920e4a300006e",
"objective_field": "000008",
"excluded_fields": ["000009"],
"objective_weights": [["yes", 10], ["no", 1]]
}'If a class is not listed in the objective_weights, it is assumed to have a weight of 1. This means the example below is equivalent to the example above.
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/52bcdc513c1920e4a300006e",
"objective_field": "000008",
"excluded_fields": ["000009"],
"objective_weights": [["yes", 10]]
}'Weights of zero are valid as long as there are some positive valued weights. If every weight does end up zero (this is possible with sampled datasets), then the resulting model will have a single node with a nil output.
Automatic Balancing
Finally, we provide a convenience shortcut for specifying weights for a classification objective which are proportional to their category counts, by means of the balance_objective flag.
For instance, if the category counts of the objective field are, say:
[["Iris-versicolor", 20], ["Iris-virginica", 10], ["Iris-setosa", 5]]the request:
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/52bb141a3c19200bdf00000c",
"balance_objective": true
}'would be equivalent to:
curl "https://bigml.io/model?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/52bb141a3c19200bdf00000c",
"objective_weights": [
["Iris-versicolor", 1],
["Iris-virginica", 2],
["Iris-setosa", 4]]}'The next table summarizes all the available arguments to use weights.
The nodes for a weighted tree will include a weight and weighted_objective_distribution, which are the weighted analogs of count and objective_distribution. Confidence, importance, and pruning calculations also take weights into account.
{
"id":0,
"children":[
{
"id":1,
"children":[
{
"output":"Iris-virginica",
"count":10,
"objective_summary":{
"categories":[
[
"Iris-virginica",
10
]
]
},
"predicate":{
"value":1.7,
"operator":">",
"field":"000003"
},
"weighted_objective_summary":{
"categories":[
[
"Iris-virginica",
10
]
]
},
"weight":10,
"confidence":0.72246,
"id":2
},
{
"output":"Iris-versicolor",
"count":20,
"objective_summary":{
"categories":[
[
"Iris-versicolor",
20
]
]
},
"predicate":{
"value":1.7,
"operator":"<=",
"field":"000003"
},
"weighted_objective_summary":{
"categories":[
[
"Iris-versicolor",
20
]
]
},
"weight":20,
"confidence":0.83887,
"id":3
}
],
"weighted_objective_summary":{
"categories":[
[
"Iris-versicolor",
20
],
[
"Iris-virginica",
10
]
]
},
"weight":30,
"predicate":{
"value":0.6,
"operator":">",
"field":"000003"
},
"confidence":0.4878,
"count":30,
"output":"Iris-versicolor",
"objective_summary":{
"categories":[
[
"Iris-versicolor",
20
],
[
"Iris-virginica",
10
]
]
}
},
{
"output":"Iris-setosa",
"count":5,
"objective_summary":{
"categories":[
[
"Iris-setosa",
5
]
]
},
"predicate":{
"value":0.6,
"operator":"<=",
"field":"000003"
},
"weighted_objective_summary":{
"categories":[
[
"Iris-setosa",
100
]
]
},
"weight":100,
"confidence":0.56551,
"id":4
}
],
"weighted_objective_summary":{
"categories":[
[
"Iris-setosa",
100
],
[
"Iris-versicolor",
20
],
[
"Iris-virginica",
10
]
]
},
"weight":130,
"predicate":true,
"confidence":0.60745,
"count":35,
"output":"Iris-setosa",
"objective_summary":{
"categories":[
[
"Iris-versicolor",
20
],
[
"Iris-virginica",
10
],
[
"Iris-setosa",
5
]
]
}
}
Ensembles
Last Updated: Thursday, 2018-10-18 13:19
Depending on the nature of your data and the specific parameters of the ensemble, you can significantly boost the predictive performance for single models, using exactly the same data.
You can create an ensemble just as you would create a model with the following three basic machine learning techniques: bagging, random decision forests, and gradient tree boosting.
Bagging, also known as bootstrap aggregating, is one of the simplest ensemble-based strategies but often outperforms strategies that are more complex. The basic idea is to use a different random subset of the original dataset for each model in the ensemble. Specifically BigML uses by default a sampling rate of 100% with replacement for each model. You can read more about bagging here.
Random decision forests is the second ensemble-based strategy that BigML provides. It consists, essentially, in selecting a new random set of the input fields at each split while an individual model is being built instead of considering all the input fields. To create a random decision forest you just need to set the randomize argument to true. You can read more about random decision forests here.
Gradient tree boosting is the third strategy whose predictions are additive. Each tree modifies the predictions of the previously grown tree. You must specify the boosting argument in order to apply this technique.
BigML.io allows you to create, retrieve, update, delete your ensemble. You can also list all of your ensembles.
Jump to:
- Ensemble Base URL
- Creating an Ensemble
- Ensemble Arguments
- Gradient Boosting
- PMML
- Retrieving an Ensemble
- Ensemble Properties
- Filtering and Paginating Fields from an Ensemble
- Updating an Ensemble
- Deleting an Ensemble
- Listing Ensembles
Ensemble Base URL
You can use the following base URL to create, retrieve, update, and delete ensembles. https://bigml.io/ensembleAll requests to manage your ensembles must use HTTPS and be authenticated using your username and API key to verify your identity. See this section for more details.
Creating an Ensemble
To create a new ensemble, you need to POST to the ensemble base URL an object containing at least the dataset/id that you want to use to create the ensemble. The content-type must always be "application/json".
POST /ensemble?username=alfred;api_key=79138a622755a2383660347f895444b1eb927730; HTTP/1.1
Host: bigml.io
Content-Type: application/json
curl "https://bigml.io/ensemble?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/50e8d4f03c19202d91000004"}' BigML.io will return the newly created ensemble if the request succeeded.
{
"category": 0,
"code": 201,
"columns": 9,
"created": "2013-01-11T00:04:20.202976",
"credits": 0.7992858886718751,
"credits_per_prediction": 0.0,
"dataset": "dataset/50e8d4f03c19202d91000004",
"dataset_status": true,
"description": "",
"ensemble_sample": {
"rate": 0.8,
"replacement": true,
"seed": "my ensemble sample seed"
},
"error_models": 0,
"finished_models": 0,
"locale": "en-US",
"max_columns": 9,
"max_rows": 768,
"missing_splits": true,
"models": [
"model/50ef57043c19208c50000026",
"model/50ef57043c19208c50000029",
"model/50ef57043c19208c5000002c",
"model/50ef57053c19208c5000002f",
"model/50ef57053c19208c50000032",
"model/50ef57053c19208c50000035",
"model/50ef57063c19208c50000038",
"model/50ef57063c19208c5000003b",
"model/50ef57063c19208c5000003e",
"model/50ef57073c19208c50000041"
],
"name": "diabetes' dataset ensemble",
"number_of_evaluations": 0,
"number_of_models": 10,
"number_of_predictions": 0,
"number_of_public_predictions": 0,
"ordering": 0,
"out_of_bag": false,
"price": 0.0,
"private": true,
"project": null,
"randomize": false,
"range": [
1,
768
],
"replacement": false,
"resource": "ensemble/50ef57043c19208c50000022",
"rows": 768,
"sample_rate": 0.8,
"seed": "a0f717f2b3954111b27fcc23f5a85787",
"size": 209528,
"source": "source/50e8d4ea3c19202d91000000",
"source_status": true,
"status": {
"code": 3,
"message": "The ensemble creation has been started"
},
"tags": [
"diabetes"
],
"updated": "2013-01-11T00:04:20.203007",
"views": 0,
"white_box": false
}
Ensemble Arguments
In addition to the dataset, you can also POST the following arguments, and like models, you can use weights to deal with imbalanced datasets. Click here to find more information about weights.
| Argument | Type | Description |
|---|---|---|
|
balance_objective
optional |
Boolean, default is false |
Whether to balance classes proportionally to their category counts or not. For more information, see the Automatic Balancing section.
Example: true |
|
boosting
optional |
Object |
Gradient boosting options for the ensemble. Required to created an ensemble with boosted trees. See the Gradient Boosting section for more information.
Example: |
|
category
optional |
Integer, default is the category of the dataset |
The category that best describes the ensemble. See the category codes for the complete list of categories.
Example: 1 |
| dataset | String |
A valid dataset/id.
Example: dataset/4f66a80803ce8940c5000006 |
|
depth_threshold
optional |
Integer, default is 512 |
When the depth in the tree exceeds this value, the tree stops growing. It has no effect if it's bigger than the node_threshold.
Example: 128 |
|
description
optional |
String |
A description of the ensemble up to 8192 characters long.
Example: "This is a description of my new ensemble" |
|
ensemble_sample
optional |
Object |
The sampling to be used for each tree in the ensemble. It can contain a rate (default 1), and replacement (default true), and seed parameters. Note that this is different from the sample_rate, replacement, and seed used in other models, predictions or datasets, where sampling is applied once to the input dataset; rather, it's applied multiple times to the input, in order to create separate samplings for each tree composing the final ensemble. So there is not out_of_bag parameter here, and the seed is used to create a different seed for each of the generated trees.
Example: |
|
excluded_fields
optional |
Array, default is [], an empty list. None of the fields in the dataset is excluded. |
Specifies the fields that won't be included in the models of the ensemble
Example: |
|
fields
optional |
Object, default is {}, an empty dictionary. That is, no names or preferred statuses are changed. |
This can be used to change the names of the fields in the ensemble with respect to the original names in the dataset or to tell BigML that certain fields should be preferred. An entry keyed with the field id generated in the source for each field that you want the name updated.
Example: |
|
input_fields
optional |
Array, default is []. All the fields in the dataset |
Specifies the fields to be considered to create the ensemble.
Example: |
|
max_training_time
optional |
Integer, default is 1800 |
The maximum training time allowed for the optimization, in seconds, as a strictly positive integer. Applicable only when optimize is set to true.
Example: 3600 |
|
missing_splits
optional |
Boolean, default is true |
Defines whether to explicitly include missing field values when choosing a split while growing the models of an ensemble. When this option is enabled, in each model, generates predicates whose operators include an asterisk, such as >*, <=*, =*, or !=*. The presence of an asterisk means "or missing". So a split with the operator >* and the value 8 can be read as "x > 8 or x is missing". When using missing_splits there may also be predicates with operators = or !=, but with a null value. This means "x is missing" and "x is not missing" respectively.
Example: flase |
|
name
optional |
String, default is dataset's name |
The name you want to give to the new ensemble.
Example: "my new ensemble" |
|
node_threshold
optional |
Integer, default is 512 |
When the number of nodes in the tree exceeds this value, the tree stops growing.
Example: 1000 |
|
number_of_model_candidates
optional |
Integer, default is 128 |
The number of model candidates evaluated over the course of the optimization. Applicable only when optimize is set to true. Maximum 200 candidates.
Example: 100 |
|
number_of_models
optional |
Integer, default is 10 |
The number of models to build the ensemble. This parameter is ignored for boosted trees. See the Gradient Boosting section for more information.
Example: 100 |
|
objective_field
optional |
String, default is dataset's pre-defined objective field |
Specifies the id of the field that the ensemble will predict.
Example: "000003" |
|
objective_weights
optional |
Array |
A list of category and weight pairs. One per objective class. For more information, see the Objective Weights section.
Example: |
|
optimize
optional |
Boolean, default is false |
Whether the ensemble should be built with the automatic optimization. When it is set to true, only the following modeling properties are applied: input_fields, excluded_fields, default_numeric_value, missing_numerics, sample_rate, max_training_time, objective_weights, weight_field, and number_of_model_candidates
Example: true |
|
ordering
optional |
Integer, default is 0 (deterministic). |
Specifies the type of ordering followed to build the models of the ensemble. There are three different types that you can specify:
Example: 1 |
|
out_of_bag
optional |
Boolean, default is false |
Setting this parameter to true will return a sequence of the out-of-bag instances instead of the sampled instances. See the Section on Sampling for more details.
Example: true |
|
project
optional |
String |
The project/id you want the ensemble to belong to.
Example: "project/54d98718f0a5ea0b16000000" |
|
random_candidate_ratio
optional |
Float |
A real number between 0 and 1. When randomize is true and random_candidate_ratio is given, BigML randomizes the trees and uses random_candidate_ratio * total fields (counting the number of terms in text fields as fields). To get the final number of candidate fields we round down to the nearest integer, but if the result is 0 we'll use 1 instead. If both random_candidates and random_candidate_ratio are given, BigML ignores random_candidate_ratio.
Example: 0.2 |
|
random_candidates
optional |
Integer, default is the square root of the total number of input fields. |
Sets the number of random fields considered when randomize is true.
Example: 10 |
|
randomize
optional |
Boolean, default is false |
Setting this parameter to true will consider only a subset of the possible fields when choosing a split. See the Section on Random Decision Forests for further details.
Example: true |
|
range
optional |
Array, default is [1, max rows in the dataset] |
The range of successive instances to build the ensemble.
Example: [1, 150] |
|
replacement
optional |
Boolean, default is false |
Whether sampling should be performed with or without replacement. See the Section on Sampling for more details.
Example: true |
|
sample_rate
optional |
Float, default is 1.0 |
A real number between 0 and 1 specifying the sample rate. See the Section on Sampling for more details.
Example: 0.5 |
|
seed
optional |
String |
A string to be hashed to generate deterministic samples. See the Section on Sampling for more details.
Example: "MySample" |
|
split_candidates
optional |
Integer, default is 32 |
The number of split points that are considered whenever the tree evaluates a numeric field. Minimum 1 and maximum 1024
Example: 128 |
|
stat_pruning
optional |
Boolean, default is false |
Activates statistical pruning on each decision tree model. It doesn't apply to boosted trees.
Example: true |
|
support_threshold
optional |
Float, default is 0 |
The parameter controls the minimum amount of support each child node must contain to be valid as a possible split. So, if it is 3, then a both children of a new split must have 3 instances supporting them. Since instances may have non-integer weights, non-integer values are valid.
Example: 16 |
|
tags
optional |
Array of Strings |
A list of strings that help classify and index your ensemble.
Example: ["best customers", "2018"] |
|
webhook
optional |
Object |
A webhook url and an optional secret phrase. See the Section on Webhooks for more details.
Example: |
|
weight_field
optional |
String |
Any numeric field with no negative or missing values is valid as a weight field. Each instance will be weighted individually according to the weight field's value. For more information, see the Weight Field section.
Example: "000005" |
You can use curl to customize a new ensemble from the command line. For example, to create a new ensemble named "my ensemble", with only certain rows, and with only three fields:
curl "https://bigml.io/ensemble?$BIGML_AUTH" \
-X POST \
-H 'content-type: application/json' \
-d '{"dataset": "dataset/4f66a80803ce8940c5000006", "input_fields": ["000001", "000003"], "name": "my dataset", "range": [25, 125]}'If you do not specify a name, the dataset's name will be assigned to the new ensemble. If you do not specify a range of instances, the complete set of instances in the dataset will be used. If you do not specify any input fields, all the preferred input fields in the dataset will be included, and if you do not specify an objective field, the last field in your dataset will be considered the objective field.
Gradient Boosting
When doing boosting, the number_of_models parameter described above is no longer valid as an input. The number_of_models will now indicate the maximum number of boosting iterations explained below. Note that when gradient boosting option is applied to classification models, the actual number of models created will be a product of the number of classes (categories) and the iterations. For example, if you set boosting iterations to 12 and the number of classes is 3, then the number of models created will be 36 or less depending on whether an early stopping strategy is used or not.
In addition, our implementation of boosted trees support the following parameters, which are all part of the boosting object:
If the boosted trees are using one of the early stopping tests (early_out_of_bag or early_holdout), then it will also have a list of scores indicating the quality of the boosted trees after each iteration.
Individual trees in the boosted trees differ from trees in bagged or random forest ensembles. Primarily the difference is that boosted trees do not try to predict the objective field directly. Instead, they try to fit a gradient (correcting for mistakes made in previous iterations), and this will be stored under a new field, named gradient.
This means the predictions from boosted trees cannot be combined with using the regular ensemble combiners. Instead, boosted trees use their own combiner that relies on a few new parameters included with individual boosted trees. These new parameters will be contained in the boosting attribute in each boosted tree, which may contain the following properties.
- objective_class will indicate the class that each tree helps predict if boosting is used for a classification problem (there will be one tree for each class for every boosting iteration).
- objective_field: contains the field id of the original objective field, as boosted trees will always be regression trees whose new objective is a new generated field (the previously mentioned gradient).
- weight: captures the influence each tree has when computing predictions.
- lambda: helps regulate the strength of a tree's output. It's included for generating predictions when encountering missing data and using the proportional strategy.
Nodes in boosted trees will also contain two new boosting related parameters, g_sum and h_sum. These are sums of the first and second order gradients, and are needed for generating predictions when encountering missing data and using the proportional strategy.
For regression problems, a prediction is generated by finding the prediction from each individual tree and doing a weighted sum using each tree's weight. Predictions for classification problems are similar, but separate weighted sums are found for each objective_class. That vector of weighted sums is then transformed into class probabilities using the soft max function.
Retrieving an Ensemble
Each ensemble has a unique identifier in the form "ensemble/id" where id is a string of 24 alpha-numeric characters that you can use to retrieve the ensemble.
To retrieve an ensemble with curl:
curl "https://bigml.io/ensemble/50ef57043c19208c50000022?$BIGML_AUTH"You can also use your browser to visualize the ensemble using the full BigML.io URL or pasting the ensemble/id into the BigML.com dashboard.
Ensemble Properties
Once an ensemble has been successfully created it will have the following properties.
| Property | Type | Description |
|---|---|---|
| boosting | Object | Gradient boosting options for the ensemble including scores which indicates the quality of the boosted trees after each iteration and final_iterations.See the Gradient Boosting section for more information. |
|
category
filterable, sortable, updatable |
Integer | One of the categories in the table of categories that help classify this resource according to the domain of application. |
| code | Integer | HTTP status code. This will be 201 upon successful creation of the ensemble and 200 afterwards. Make sure that you check the code that comes with the status attribute to make sure that the ensemble creation has been completed without errors. |
|
columns
filterable, sortable |
Integer | The number of fields in the ensemble. |
|
composites
filterable, sortable |
Array of Strings | The list of composite ids that reference this model. |
|
created
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the ensemble was created with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
|
credits
filterable, sortable |
Float | The number of credits it cost you to create this ensemble. |
|
credits_per_prediction
filterable, sortable, updatable |
Float | This is the number of credits that other users will consume to make a prediction with your ensemble in case you decide to make it public. |
|
dataset
filterable, sortable |
String | The dataset/id that was used to build the ensemble. |
|
dataset_status
filterable, sortable |
Boolean | Whether the dataset is still available or has been deleted. |
|
description
updatable |
String | A text describing the ensemble. It can contain restricted markdown to decorate the text. |
| distributions | Array | Unordered list of distributions for each model in the ensemble. Each distribution is an Object with a entry for the distribution of instances in the training set and the distribution of predictions in the model. See a model distribution field for more details. Note that distributions must be accessed by the model_order below. |
| ensemble_sample | Object | The sampling to be used for each tree in the ensemble. |
|
error_models
filterable, sortable |
Integer | The number of models in the ensemble that have failed. |
| excluded_fields | Array | The list of fields's ids that were excluded to build the ensemble. |
| fields_meta | Object | A dictionary with meta information about the fields dictionary. It specifies the total number of fields, the current offset, and limit, and the number of fields (count) returned. |
|
finished_models
filterable, sortable |
Integer | The number of models in the ensemble that have finished correctly. |
|
fusions
filterable, sortable |
Array of Strings | The list of fusion ids that reference this model. |
| importance | Object | Average importance per field |
| input_fields | Array | The list of input fields' ids used to build the models of the ensemble. |
| locale | String | The dataset's locale. |
|
max_columns
filterable, sortable |
Integer | The total number of fields in the dataset used to build the ensemble. |
|
max_rows
filterable, sortable |
Integer | The maximum number of instances in the dataset that can be used to build the ensemble. |
| max_training_time | Integer | The maximum training time allowed for the optimization, in seconds. |
|
missing_splits
filterable, sortable |
Boolean | Whether to explicitly include missing field values when choosing a split while growing the models of an ensemble. |
| model_order | Array | Order in which each model in the list of models was finished. The distributions above must be accessed following this index. |
| models | Array | Unordered list of model/ids that compose the ensemble. Models are ordered by the model_order above. |
|
name
filterable, sortable, updatable |
String | The name of the ensemble as your provided or based on the name of the dataset by default. |
|
node_threshold
filterable, sortable |
String | The maximum number of nodes that the model will grow. |
|
number_of_batchpredictions
filterable, sortable |
Integer | The current number of batch predictions that use this ensemble. |
|
number_of_evaluations
filterable, sortable |
Integer | The current number of evaluations that use this ensemble. |
| number_of_model_candidates | Integer | The number of model candidates evaluated over the course of the optimization. |
|
number_of_models
filterable, sortable |
Integer | The number of models in the ensemble. |
|
number_of_predictions
filterable, sortable |
Integer | The current number of predictions that use this ensemble. |
|
number_of_public_predictions
filterable, sortable |
Integer | The current number of public predictions that use this ensemble. |
| objective_field | String |
Specifies the id of the field that the ensemble predicts.
Example: "000003" |
| optimize | Boolean | Whether the ensemble was built with the automatic optimization. |
|
optiml
filterable, sortable |
String | The optiml/id that created this ensemble. |
|
optiml_status
filterable, sortable |
Boolean | Whether the OptiML is still available or has been deleted. |
|
ordering
filterable, sortable |
Integer |
The order used to chose instances from the dataset to build the models of the ensemble. There are three different types:
|
|
out_of_bag
filterable, sortable |
Boolean | Whether the out-of-bag instances were used to create the ensemble instead of the sampled instances. |
|
price
filterable, sortable, updatable |
Float | The price other users must pay to clone your ensemble. |
|
private
filterable, sortable, updatable |
Boolean | Whether the ensemble is public or not. |
|
project
filterable, sortable, updatable |
String | The project/id the resource belongs to. |
|
random_candidate_ratio
filterable, sortable |
Float | The random candidate ratio considered when randomize is true. |
|
random_candidates
filterable, sortable |
Integer | The number of random fields considered when randomize is true. |
|
randomize
filterable, sortable |
Boolean | Whether the splits of each model in the ensemble considered only a random subset of the fields or all the fields available. |
| range | Array | The range of instances used to build the models of the ensemble. |
|
replacement
filterable, sortable |
Boolean | Whether the instances sampled to build the ensemble were selected using replacement or not. |
| resource | String | The ensemble/id. |
|
rows
filterable, sortable |
Integer | The total number of instances used to build the models of the ensemble |
|
sample_rate
filterable, sortable |
Float | The sample rate used to select instances from the dataset to build the models of the ensemble. |
|
seed
filterable, sortable |
String | The string that was used to generate the sample. |
|
shared
filterable, sortable, updatable |
Boolean | Whether the ensemble is shared using a private link or not. |
| shared_hash | String | The hash that gives access to this ensemble if it has been shared using a private link. |
| sharing_key | String | The alternative key that gives read access to this ensemble. |
|
size
filterable, sortable |
Integer | The number of bytes of the dataset that were used to create this ensemble. |
|
source
filterable, sortable |
String | The source/id that was used to build the dataset. |
|
source_status
filterable, sortable |
Boolean | Whether the source is still available or has been deleted. |
|
split_candidates
filterable, sortable |
Integer | The number of split points that are considered whenever the tree evaluates a numeric field. Minimum 1 and maximum 1024 |
| status | Object | A description of the status of the ensemble. It includes a code, a message, and some extra information. See the table below. |
|
subscription
filterable, sortable |
Boolean | Whether the ensemble was created using a subscription plan or not. |
|
support_threshold
filterable, sortable |
Float | The parameter controls the minimum amount of support each child node must contain to be valid as a possible split. |
|
tags
filterable, updatable |
Array of Strings | A list of user tags that can help classify and index this resource. |
|
updated
filterable, sortable |
ISO-8601 Datetime. | This is the date and time in which the ensemble was updated with microsecond precision. It follows this pattern yyyy-MM-ddThh:mm:ss.SSSSSS. All times are provided in Coordinated Universal Time (UTC). |
| webhook | Object | A webhook url and an optional secret phrase. See the Section on Webhooks for more details. |
|
white_box
filterable, sortable |
Boolean | Whether the ensemble is publicly shared as a white-box. |
Ensemble Status
Creating a ensemble is a process that can take just a few seconds or a few days depending on the size of the dataset used as input, the number of models, and on the workload of BigML's systems. The ensemble goes through a number of states until its fully completed. Through the status field in the ensemble you can determine when the ensemble has been fully processed and ready to be used to create predictions. These are the properties that an ensemble's status has:
| Property | Type | Description |
|---|---|---|
| code | Integer | A status code that reflects the status of the ensemble creation. It can be any of those that are explained here. |
| elapsed | Integer | Number of milliseconds that BigML.io took to process the ensemble. |
| message | String | A human readable message explaining the status. |
| progress | Float, between 0 and 1 | How far BigML.io has progressed building the ensemble. |
Once a ensemble has been successfully created, it will look like:
{
"balance_objective":false,
"category":0,
"code":200,
"columns":9,
"created":"2016-07-08T18:26:36.351000",
"credits":0.09991455078125,
"credits_per_prediction":0,
"dataset":"dataset/5747ae334e172785fd000000",
"dataset_field_types":{
"categorical":1,
"datetime":0,
"effective_fields":9,
"items":0,
"numeric":8,
"preferred":9,
"text":0,
"total":9
},
"dataset_status":true,
"datasets":[],
"description":"",
"ensemble_sample": {
"rate": 0.8,
"replacement": true,
"seed": "my ensemble sample seed"
},
"distributions":[
{
"importance":[
[
"000001",
0.35199
],
[
"000005",
0.21031
],
[
"000006",
0.13889
],
[
"000007",
0.11932
],
[
"000002",
0.08733
],
[
"000003",
0.03628
],
[
"000000",
0.03297
],
[
"000004",
0.02291
]
],
"predictions":{
"categories":[
[
"false",
512
],
[
"true",
256
]
]
},
"training":{
"categories":[
[
"false",
512
],
[
"true",
256
]
]
}
},
{
"importance":[
[
"000001",
0.33276
],
[
"000005",
0.24432
],
[
"000006",
0.15996
],
[
"000007",
0.15378
],
[
"000003",
0.05712
],
[
"000002",
0.02431
],
[
"000000",
0.02199
],
[
"000004",
0.00575
]
],
"predictions":{
"categories":[
[
"false",
515
],
[
"true",
253
]
]
},
"training":{
"categories":[
[
"false",
515
],
[
"true",
253
]
]
}
},
{
"importance":[
[
"000001",
0.34203
],
[
"000005",
0.21501
],
[
"000006",
0.15173
],
[
"000007",
0.11991
],
[
"000003",
0.05734
],