Preprocessing splitters groupkfold

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.GroupKFold

Cuts folds that never split a group, as GroupKFold does.

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

Parameters — groups is one group label per row, any integers. foldCount is how many folds to cut, at least two and at most the number of distinct groups. randomState is scikit-learn's random_state, in [0, 2³² − 1]: it selects GroupKFold(shuffle=True).

Returns — one FoldSplit per fold, in order, each index list ascending. Every row of a group is held out by the same fold.

Exceptions — ArgumentOutOfRangeException when there are fewer than two rows, when foldCount is below two or above the row count, or when randomState is outside numpy's range. ArgumentException when foldCount is above the number of distinct groups, which the reference refuses too.

Example — four groups of three, two, one and four rows, over two folds.

using Lodestar.Preprocessing;

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

// The largest group first, each to the fold with fewest rows so far: 4 → fold 0, 1 → fold 1, 2 → 1, 3 → 0.
IReadOnlyList<FoldSplit> folds = Splitters.GroupKFold(groups, foldCount: 2);
string first = string.Join(",", folds[0].TestIndices);   // => 5,6,7,8,9
string second = string.Join(",", folds[1].TestIndices);  // => 0,1,2,3,4

// Shuffled with scikit-learn's own seed: the distinct groups are permuted, then cut into runs.
string seeded = string.Join(",", Splitters.GroupKFold(groups, 2, randomState: 0)[0].TestIndices);  // => 5,6,7,8,9

Remarks — unshuffled, the folds balance rows: groups go largest first, each to the fold holding the fewest rows, the first such fold on a tie. Groups of equal size go in descending label order, which is what the reference's reversed stable sort gives. Shuffled, they balance groups: the distinct groups, ascending, are permuted by numpy's generator and cut by np.array_split, the first groupCount % foldCount runs one group longer, whatever their sizes.

Both are replayed index for index from tests/oracles/preprocessing_splitters_seeded.json.

Applies to — net10.0, netstandard2.0.

See also — Splitters.StratifiedGroupKFold, Splitters.KFold, FoldSplit.