Cluster classification

Cluster classification | BigML.com

FREE

Cluster Classification

Details:

Created	Mon, 1 Aug 2016 18:02:56 +0000
Published	Fri, 12 Aug 2016 02:30:08 +0000

Description:

Find the global field importance across a cluster

Please see the readme for more information.

Tags:

Clusters Importance

URL:

https://bigml.com/user/akashenfelter/gallery/script/579f8ed0af447f2234000ddb

Script FREE

Source code

;; Given a cluster, creates a batch centroid and uses the output
;; dataset where each instance is labeled by centroid to create an
;; ensemble (random forest), and returns the importance of each field
;; averaged over the ensemble.

;; Given a cluster, returns a batchcentroid id created from that
;; cluster whose output dataset is the original cluster dataset
;; appended with the centroids the instances are part of.

(define (label-by-cluster cl-id) 
  (let (cl (fetch cl-id)
        cl-ds (cl "dataset")
        cl-fields (cl "input_fields"))
    (create-and-wait-batchcentroid
     {"cluster" cl-id 
      "dataset" cl-ds 
      "output_dataset" true
      "output_fields" cl-fields})))

;; Given a list of pairs, creates the equivalent map

(define (list-to-map nested-list)
  (let (key-list (map head nested-list)
        value-list (map (lambda (x) (last x)) nested-list))
    (make-map key-list value-list)))

;; Given a function of two values and two maps, returns a new map
;; where if a key appears in only one of the original maps its value
;; is unchanged, but if it appears in both its value is the function
;; of the two original values.  

(define (merge-with fn map1 map2) 
  (let (subset-keys (filter (lambda (x) (contains? map1 x)) (keys map2))
        subset-values (map (lambda (x) 
                             (fn (get map1 x) (get map2 x))) subset-keys) 
        subset (make-map subset-keys subset-values)) 
    (merge map1 (merge map2 subset))))

;; Given a ensemble id and a list of field ids, returns a map of those
;; field ids and average importances

(define (make-importance-map en-id ids)
  (let (models ((fetch en-id) ["models"])
        len-models (count models)
        imp-lst (map (lambda (md-id) 
                       ((fetch md-id) ["model" "importance"])) 
                     models)
        imp-map (map list-to-map imp-lst)
        imp-sum (reduce (lambda (x y) (merge-with + x y)) {} imp-map))
    (make-map ids (map (lambda (x) 
                         (if (contains? imp-sum x) 
                           (/ (get imp-sum x) len-models) 0)) 
                       ids))))

;; Given a cluster-id, builds a random forest on centroid membership,
;; and returns a paired list of field importance averaged across all
;; models in the forest

(define (cluster-classification cl-id)
  (let (bc-id (label-by-cluster cl-id)
        bc (fetch bc-id)
        old-ids (bc "output_fields")
        labeled (bc "output_dataset_resource")
        new-ids (butlast ((fetch labeled) "input_fields"))
        cl-en (create-and-wait-ensemble labeled {"randomize" true})
        ds-fields ((fetch labeled) "fields")
        imp-map (make-importance-map cl-en new-ids)
        name-map (make-map new-ids 
                           (map (lambda (x) 
                                  (ds-fields [x "name"])) 
                                new-ids))
        id-map (make-map new-ids old-ids)       
        total-map (map (lambda (x) 
                         (make-map ["field id", "name", "importance"] 
                                   [(get id-map x), 
                                    (get name-map x), 
                                    (get imp-map x)])) 
                       (keys imp-map)))
  (delete* [labeled bc-id cl-en])
  (reverse (sort-by-key "importance" total-map))))

;; Putting it all together...
(define importance-list 
  (cluster-classification cluster-id))

Description

Inputs

Outputs

License: Creative Commons Zero V1.0
Script FREE

Cluster classification | BigML.com

Sending Request...

COMPANY

PRODUCT

BUSINESS

TRAINING

GALLERY

License

Embed this Script in your web site

COMPANY

PRODUCT

BUSINESS

TRAINING

GALLERY