Preprocessing splitters stratifiedtraintest

Development build. This page describes main, not a released package. The latest published Lodestar.Preprocessing is 0.2.0 — read its documentation.

Home › Preprocessing › Splitting

Splitters.StratifiedTrainTest

Holds out a share that keeps each label's proportion, as train_test_split(stratify=y) does.

public static TrainTestSplit StratifiedTrainTest(ReadOnlySpan<int> labels, double testFraction, long randomState)

Parameters — labels is one class label per row; every class needs at least two rows. testFraction is the share to hold out, strictly inside (0, 1). randomState is scikit-learn's random_state, in [0, 2³² − 1].

Returns — a TrainTestSplit of ceil(n · testFraction) test rows and the rest for training, each index list ascending.

Exceptions — ArgumentOutOfRangeException when there are fewer than two rows, when testFraction is outside (0, 1) or rounds up to every row, or when randomState is out of range. ArgumentException, in the reference's order, when a class has a single row, or when either side would hold fewer rows than there are classes.

Example — twelve rows in three classes, a quarter held out.

using Lodestar.Preprocessing;

int[] labels = [0, 0, 0, 0, 0, 0, 1, 1, 1, 2, 2, 2];

TrainTestSplit split = Splitters.StratifiedTrainTest(labels, testFraction: 0.25, randomState: 42);

// One row of each class: 3 test rows split 6 : 3 : 3 as closely as whole rows allow.
string held = string.Join(",", split.TestIndices);   // => 3,6,11
string fit = string.Join(",", split.TrainIndices);   // => 0,1,2,4,5,7,8,9,10

Remarks — StratifiedShuffleSplit's algorithm: each class's share of the training rows, then of the test rows, from the approximate mode of the multivariate hypergeometric; then each class's rows, ascending, read through a permutation and cut there. Where two classes' remainders tie, the reference picks which one gets the extra row at random, which is why this takes the seed rather than a permutation: no permutation a caller could pass carries that choice (decision 0008).

There is no unseeded form. The reference refuses stratify with shuffle=False.

Applies to — net10.0, netstandard2.0.

See also — Splitters.TrainTest, Splitters.StratifiedKFold, TrainTestSplit.