Metrics 0.3.0 pairconfusionmatrix compute

Lodestar.Metrics 0.3.0. 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.

PairConfusionMatrix.Compute

Counts the pairs two labellings agree and disagree about.

public static PairConfusionMatrix Compute(ReadOnlySpan<int> labelsTrue, ReadOnlySpan<int> labelsPred)

Parameters — labelsTrue is the reference labelling and labelsPred the one being scored, one label per sample and the same length. The label values carry no meaning: only which samples share one does.

Returns — PairConfusionMatrix, four long counts that sum to the square of the sample count.

Exceptions — ArgumentException when the two labellings disagree in length. An empty input is not an error: every count is 0.

Example — verified against the two worked examples in scikit-learn's own docstring.

using Lodestar.Metrics;

PairConfusionMatrix perfect = PairConfusionMatrix.Compute([0, 0, 1, 1], [1, 1, 0, 0]);
long same = perfect.SameInBoth;   // => 4

PairConfusionMatrix penalised = PairConfusionMatrix.Compute([0, 0, 1, 2], [0, 0, 1, 1]);
long overMerged = penalised.SameInPredictedOnly;   // => 2

Remarks — the first example is a renaming: [1,1,0,0] is [0,0,1,1] with the two labels swapped, so every pair that is together in one labelling is together in the other, and every pair apart in one is apart in the other — DifferentInBoth is 8, SameInBoth is 4, and the other two quadrants are 0. Renaming a cluster changes nothing this type counts, which is exactly what RandIndex.Score needs to stay 1 on it.

The second is the penalised case scikit-learn's own documentation uses: the prediction merges two of the truth's clusters into one. SameInPredictedOnly is 2 — two pairs the prediction now calls together that the truth kept apart — where a clean split would have left it 0.

Applies to — net10.0, netstandard2.0.

See also — PairConfusionMatrix, PairConfusionMatrix.ToArray, RandIndex.Score.