Metrics logloss multiclass

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

LogLoss.MultiClass

The cross-entropy over a probability matrix — sklearn.metrics.log_loss with 2-D probabilities.

public static double MultiClass(ReadOnlySpan<int> yTrue, ReadOnlySpan<double> yProba, int classCount, bool normalize = true, ReadOnlySpan<double> sampleWeight = default)

Parameters — yTrue is the true class index of each sample, in [0, classCount). yProba is the class probabilities row-major: sample 0's classes, then sample 1's. classCount is how many classes each row scores. normalize divides by the total weight; pass false for the sum. sampleWeight is one weight per sample, or empty.

Returns — double, 0 or above and unbounded. Only the column of the true class contributes, so the score depends on the other columns solely through whatever normalisation the caller applied.

Exceptions — ArgumentException when yProba is not yTrue.Length × classCount, when a label is not a class index below classCount, or when a probability falls outside [0, 1]. ArgumentOutOfRangeException when classCount is below two.

Example — four samples over three classes.

using Lodestar.Metrics;

int[] truth = [0, 1, 2, 1];
double[] probabilities =
[
    0.7, 0.2, 0.1,
    0.1, 0.8, 0.1,
    0.2, 0.2, 0.6,
    0.3, 0.4, 0.3,
];

double loss = LogLoss.MultiClass(truth, probabilities, classCount: 3);  // => 0.5017…

Remarks — a row that does not sum to 1 is neither refused nor renormalised. The reference warns and scores the values as given; there is no warning channel here, so the number is the only signal — measured, halving every row above takes the loss to 1.1948…. This is the one place where RocAuc.MultiClass is stricter than its own reference and this is not: that one refuses a row that does not sum to 1.

Applies to — net10.0, netstandard2.0.

See also — LogLoss.Score, BrierScore.MultiClass, the Python equivalence table.