Metrics 0.3.0 randindex score

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.

RandIndex.Score

How many pairs of samples the two partitions treat the same way, over all pairs.

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, 1]. 1 when the two labellings agree on every pair, whichever way; 0 only when they agree on none.

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

Example — this is AdjustedRand.Score before the correction, and the gap between the two is the correction made legible.

using Lodestar.Metrics;

int[] truth = [0, 0, 0, 1, 1, 1];
int[] split = [0, 0, 1, 2, 2, 2];

double rand = RandIndex.Score(truth, split);            // => 0.8666666666666667
double adjusted = AdjustedRand.Score(truth, split);     // => 0.7058823529411765

Remarks — 0.867 against 0.706 is the pair of numbers that makes the correction concrete: most pairs already agree by construction on a small sample, so an uncorrected score stays high even for a clustering that is only partly right. AdjustedRand.Score subtracts what agreement by chance alone would already give, which is why it reads lower on the same input.

Because it is uncorrected, two independent labellings score well above zero here. [0,0,1,1] against [0,1,0,1] scores 0.333 under this metric and -0.5 under AdjustedRand.Score — the same disagreement, read two different ways. That makes this the wrong choice for comparing clusterings that use different numbers of clusters, where chance agreement itself varies with cluster count.

An empty input and a single sample both score 1, as they do for every other metric in this namespace: agreeing about nothing is agreeing.

Applies to — net10.0, netstandard2.0.

See also — AdjustedRand.Score, PairConfusionMatrix.Compute, the Python equivalence table.