Development build. This page describes main, not a released package. The latest published Lodestar.Metrics is 0.4.0 — read its documentation.

Home › Metrics › Classification metrics

Both class likelihood ratios — sklearn.metrics.class_likelihood_ratios.

public static LikelihoodRatios Compute(ReadOnlySpan<int> yTrue, ReadOnlySpan<int> yPred, int posLabel = 1, double undefinedPositive = NaN, double undefinedNegative = NaN, ReadOnlySpan<double> sampleWeight = default)

Parameters — yTrue and yPred are the true and predicted labels, of the same length, over at most two distinct values. posLabel is the label counted as positive, 1 by default. undefinedPositive and undefinedNegative are what each ratio answers when it has no value; the defaults reproduce replace_undefined_by=nan. sampleWeight is one weight per sample, or empty.

Returns — a LikelihoodRatios carrying Positive and Negative.

Exceptions — ArgumentException when the inputs disagree in length, are empty, the weights do not match, hold a non-finite value or are zero throughout, or more than two distinct labels occur — with the reference's own sentence, "class_likelihood_ratios only supports binary classification problems." — or only one, where the reference fails unpacking a 1 × 1 matrix; and when the two do not include posLabel.

Example — six samples, half of them positive.

using Lodestar.Metrics;

int[] truth = [0, 1, 1, 0, 1, 0];
int[] predicted = [0, 1, 0, 0, 1, 1];

LikelihoodRatios ratios = LikelihoodRatios.Compute(truth, predicted);
double positive = ratios.Positive;  // => 2

A positive prediction doubles the odds; the negative ratio on the same input is 0.5, so a negative prediction halves them.

Remarks — two parameters where the reference has one. replace_undefined_by takes either a scalar or a mapping of {"LR+": …, "LR-": …}, a union C# has no equivalent of; passing the same value to both reproduces the scalar form and passing different ones reproduces the mapping.

Each ratio is computed from the counts, tp·(tn+fp) / (fp·(tp+fn)) and fn·(tn+fp) / (tn·(tp+fn)), as the reference computes it — not through 1 − specificity, which is how this example used to read 1.9999… — and each is replaced only where its own count vanishes. With no positive sample a ratio that is not replaced comes out 0/0, NaN. The type page has the table of the undefined shapes.

Applies to — net10.0, netstandard2.0.

See also — Precision.Score, Recall.Score, ConfusionMatrix.Compute, the Python equivalence table.