Metrics adjustedrand score

Development build. This page describes main, not a released package. The latest published Lodestar.Metrics is 0.3.0 — read its documentation.

AdjustedRand.Score

How many pairs of samples the two partitions agree about, minus what chance would give.

public static double Score(ReadOnlySpan<int> labelsTrue, ReadOnlySpan<int> labelsPred)

Parameters — labelsTrue is the reference partition 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 — double in [-0.5, 1]. 1 is the same partition however it is named, 0 is what a random labelling scores, and a negative number is agreement worse than chance.

Exceptions — ArgumentException when the two labellings disagree in length. An empty input is not an error: it scores 1.

Example — a clustering that got one sample wrong, against one that split everything.

using Lodestar.Metrics;

int[] truth = [0, 0, 1, 1, 2, 2];
int[] almost = [0, 0, 1, 2, 2, 2];
int[] alone = [0, 1, 2, 3, 4, 5];

double good = AdjustedRand.Score(truth, almost);   // => 0.4444…
double useless = AdjustedRand.Score(truth, alone);   // => 0

Remarks — the correction for chance is a subtraction, so this can and does go negative: [0,0,1,1] against [0,1,0,1] scores -0.5, measured. Read that as systematically disagreeing rather than as an error.

An empty input scores 1, as does a single sample, and so does one cluster covering everything when the truth is also one cluster — all three are the same partition, and scikit-learn answers them before computing anything. Against a two-cluster truth, one cluster scores 0 instead.

Applies to — net10.0, netstandard2.0.

See also — VMeasure.Score, NormalizedMutualInformation.Score, the Python equivalence table.