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

Cuts group folds that keep each label's share, as StratifiedGroupKFold does.

public static IReadOnlyList<FoldSplit> StratifiedGroupKFold(ReadOnlySpan<int> labels, ReadOnlySpan<int> groups, int foldCount)
public static IReadOnlyList<FoldSplit> StratifiedGroupKFold(ReadOnlySpan<int> labels, ReadOnlySpan<int> groups, int foldCount, long randomState)

Parameters — labels is one class label per row. groups is one group label per row, as many as labels. foldCount is how many folds to cut. randomState is scikit-learn's random_state, in [0, 2³² − 1]: it selects StratifiedGroupKFold(shuffle=True).

Returns — one FoldSplit per fold, in order, each index list ascending. No group straddles two folds.

Exceptions — ArgumentOutOfRangeException when the row count, foldCount or randomState is out of range. ArgumentException when the two spans differ in length, when foldCount is above every class's count, or when it is above the number of distinct groups. ArgumentException also when the groups or the folds times the classes are more cells than one array holds (#1314).

Example — two classes over four groups, in two folds.

using Lodestar.Preprocessing;

int[] labels = [0, 0, 1, 1, 0, 1, 0, 1, 1, 0];
int[] groups = [1, 1, 1, 2, 2, 3, 4, 4, 4, 4];

IReadOnlyList<FoldSplit> folds = Splitters.StratifiedGroupKFold(labels, groups, foldCount: 2);
string first = string.Join(",", folds[0].TestIndices);   // => 0,1,2,6,7,8,9
string second = string.Join(",", folds[1].TestIndices);  // => 3,4,5

string seeded = string.Join(",", Splitters.StratifiedGroupKFold(labels, groups, 2, randomState: 0)[0].TestIndices);  // => 5,6,7,8,9

Remarks — the reference's greedy search, step for step: groups by descending spread of their class counts, each to the fold that minimises the mean, over classes, of the spread of the folds' shares of that class; a fold within np.isclose of the best wins if it holds fewer rows. The shuffled form only changes the order ties between groups are met in.

The spreads are numpy's to the last bit, because a tie between two folds decides where a group goes. numpy sums a row pairwise from its first element and a column row by row; a random differential run found a five-class case whose fold turns on that order, and it is frozen in the corpus. A fold count above the smallest class makes the reference warn, and is accepted here.

Applies to — net10.0, netstandard2.0.

See also — Splitters.GroupKFold, Splitters.StratifiedKFold, FoldSplit.