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.

The binary hinge loss — sklearn.metrics.hinge_loss.

public static double Score(ReadOnlySpan<int> yTrue, ReadOnlySpan<double> predDecision, int posLabel = 1, ReadOnlySpan<double> sampleWeight = default)

Parameters — yTrue is the true labels, one per sample. predDecision is the decision value per sample: positive on posLabel's side of the boundary, and further from zero the more confident — not a probability. posLabel is the label on the positive side, 1 by default. sampleWeight is one weight per sample, or empty.

Returns — double, 0 or above and unbounded. 0 when every sample sits on the right side by a margin of at least 1.

Exceptions — ArgumentException when the inputs disagree in length, are empty, or the weights do not match.

Example — four samples, two of them inside the margin.

using Lodestar.Metrics;

int[] truth = [-1, 1, 1, -1];
double[] decisions = [-0.5, 1.2, 0.3, 0.8];

double loss = HingeLoss.Score(truth, decisions);  // => 0.75

The second sample is right by more than 1 and costs nothing; the fourth is on the wrong side by 0.8 and costs 1.8.

Remarks — only the sign of the decision is compared against the label, so relabelling cannot move the number. On a truth carrying one class only, scikit-learn returns a value computed against the wrong side — 1.65 where the margins say 0.35 — because its LabelBinarizer has nothing to contrast; the type page has that divergence in full.

Applies to — net10.0, netstandard2.0.

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