Stratified_shuffle_split (p.modelkit.0.5.0.doc.modelkit.Modelkit.Stratified_shuffle

Module Modelkit.Stratified_shuffle_split

Shuffled partitions with proportional class allocation; defaults to ten splits and a 10% test fraction. Requires aligned classification labels, at least two rows per class, and at least as many rows per partition as classes. Largest-remainder allocation fills training first, then test from remaining rows, with seeded tie breaking. Extreme imbalance can still omit a rare class from a partition. Classes follow first appearance order. Random streams are deterministic for the same input and seed, independent of execution scheduling; they do not reproduce NumPy random streams. Groups do not constrain these splits; use group-aware splitting when group exclusion is required.

type params = {

splits : int;
train_size : Split_size.t option;
test_size : Split_size.t option;

}

type t

val create : 
  ?splits:int ->
  ?train_size:Split_size.t ->
  ?test_size:Split_size.t ->
  unit ->
  (t, Error.t) result

include SPLITTER
  with type t := t
   and type params := params
   and type target = Target.classification Target.t
   and type rng = Rng.t

include SPECIFICATION with type t := t with type params := params

val clone : t -> t

val params : t -> params

type target = Target.classification Target.t

type rng = Rng.t

val split : 
  t ->
  rng:rng ->
  ?groups:Groups.t ->
  x:Matrix.t ->
  y:target option ->
  unit ->
  ((Row_view.t * Row_view.t) array, Error.t) result