Lodestar.Preprocessing 0.2.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

Home › Splitting

Holds out the last rows, as train_test_split(shuffle=False) does.

public static TrainTestSplit TrainTest(int sampleCount, double testFraction)
public static TrainTestSplit TrainTest(int sampleCount, double testFraction, ReadOnlySpan<int> order)

Parameters — sampleCount is how many rows there are, at least two. testFraction is the share to hold out, strictly inside (0, 1). order is a permutation of the rows whose first ceil(sampleCount · testFraction) entries are held out; an empty span holds out the last rows instead.

Returns — a TrainTestSplit carrying the training and test indices, each ascending.

Exceptions — ArgumentOutOfRangeException when sampleCount is below two, when testFraction is NaN or outside (0, 1), or when the fraction rounds up to every row and leaves nothing to fit on. ArgumentException when order is neither empty nor a permutation of the rows.

Example — a quarter of ten rows, then the same fraction read in another order.

using Lodestar.Preprocessing;

TrainTestSplit split = Splitters.TrainTest(sampleCount: 10, testFraction: 0.25);

// ceil(10 × 0.25) = 3, and they are the last three rows.
string test = string.Join(",", split.TestIndices);    // => 7,8,9
string train = string.Join(",", split.TrainIndices);  // => 0,1,2,3,4,5,6

int[] order = [9, 4, 1, 7, 0, 3, 6, 8, 2, 5];
// With an order the test rows are its first three, as ShuffleSplit takes permutation[:n_test].
string held = string.Join(",", Splitters.TrainTest(10, 0.25, order).TestIndices);  // => 1,4,9

Remarks — the held-out count is ceil(sampleCount · testFraction), which is the reference's rounding; at n = 10 and 0.25 it is three rows, not two.

The order's head is held out, the unshuffled split's tail. That asymmetry is the reference's own: train_test_split(shuffle=False) tests on the last rows, while ShuffleSplit — which train_test_split calls when it shuffles — tests on permutation[:n_test]. So passing the permutation scikit-learn drew reproduces its split, and the identity order holds out the first rows, not the same rows as the two-argument overload.

Stratification is not offered here. The reference refuses stratify with shuffle=False outright, and with a caller-supplied permutation the honest version is Splitters.StratifiedKFold with foldCount = round(1 / testFraction), taking one fold as the test set: at a quarter that is four folds, and fold 0 is the held-out rows.

Applies to — net10.0, netstandard2.0.

See also — TrainTestSplit, Splitters.KFold.