Liking cljdoc? Tell your friends :D
Clojure only.

tablecloth.api.split


splitclj

(split ds)
(split ds split-type)
(split ds split-type {:keys [seed parallel?] :as opts})

Split given dataset into 2 or more (holdout) splits

As the result two new columns are added:

  • :$split-name - with subgroup name
  • :$split-id - fold id/repetition id

split-type can be one of the following:

  • :kfold - k-fold strategy, :k defines number of folds (defaults to 5), produces k splits
  • :bootstrap - :ratio defines ratio of observations put into result (defaults to 1.0), produces 1 split
  • :holdout - split into two parts with given ratio (defaults to 2/3), produces 1 split
  • :loo - leave one out, produces the same number of splits as number of observations

:holdout can accept also probabilites or ratios and can split to more than 2 subdatasets

Additionally you can provide:

  • :seed - for random number generator
  • :repeats - repeat procedure :repeats times
  • :partition-selector - same as in group-by for stratified splitting to reflect dataset structure in splits.
  • :split-names names of subdatasets different than default, ie. [:train :test :split-2 ...]
  • :split-col-name - a column where name of split is stored, either :train or :test values (default: :$split-name)
  • :split-id-col-name - a column where id of the train/test pair is stored (default: :$split-id)

Rows are shuffled before splitting.

In case of grouped dataset each group is processed separately.

See more

Split given dataset into 2 or more (holdout) splits

As the result two new columns are added:

* `:$split-name` - with subgroup name
* `:$split-id` - fold id/repetition id

`split-type` can be one of the following:

* `:kfold` - k-fold strategy, `:k` defines number of folds (defaults to `5`), produces `k` splits
* `:bootstrap` - `:ratio` defines ratio of observations put into result (defaults to `1.0`), produces `1` split
* `:holdout` - split into two parts with given ratio (defaults to `2/3`), produces `1` split
* `:loo` - leave one out, produces the same number of splits as number of observations

`:holdout` can accept also probabilites or ratios and can split to more than 2 subdatasets

Additionally you can provide:

* `:seed` - for random number generator
* `:repeats` - repeat procedure `:repeats` times
* `:partition-selector` - same as in `group-by` for stratified splitting to reflect dataset structure in splits.
* `:split-names` names of subdatasets different than default, ie. `[:train :test :split-2 ...]`
* `:split-col-name` - a column where name of split is stored, either `:train` or `:test` values (default: `:$split-name`)
* `:split-id-col-name` - a column where id of the train/test pair is stored (default: `:$split-id`)

Rows are shuffled before splitting.

In case of grouped dataset each group is processed separately.

See [more](https://www.mitpressjournals.org/doi/pdf/10.1162/EVCO_a_00069)
sourceraw docstring

split->seqclj

(split->seq ds)
(split->seq ds split-type)
(split->seq ds
            split-type
            {:keys [split-col-name split-id-col-name]
             :or {split-col-name :$split-name split-id-col-name :$split-id}
             :as opts})

Returns split as a sequence of train/test datasets or map of sequences (grouped dataset)

Returns split as a sequence of train/test datasets or map of sequences (grouped dataset)
sourceraw docstring

cljdoc is a website building & hosting documentation for Clojure/Script libraries

× close