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 fraction of labels predicted wrongly — sklearn.metrics.hamming_loss.

public static double Score(ReadOnlySpan<int> yTrue, ReadOnlySpan<int> yPred, ReadOnlySpan<double> sampleWeight = default)
public static double Score(ReadOnlySpan<bool> yTrue, ReadOnlySpan<bool> yPred, int labelCount, ReadOnlySpan<double> sampleWeight = default)

Parameters — the first overload takes yTrue and yPred as one label per sample. The second takes them as a label matrix: one boolean per label per sample, row-major, labelCount values per row. sampleWeight is one weight per sample — per row, not per label — or empty, the default.

Returns — double in [0, 1]. 0 when everything is right.

Exceptions — ArgumentException when the inputs disagree in length, are empty, when the matrix is not a whole number of rows of labelCount, or when the weights do not match the sample count.

Example — four samples over three classes, one wrong.

using Lodestar.Metrics;

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

double loss = HammingLoss.Score(truth, predicted);  // => 0.25

Over a matrix it counts labels rather than samples, which is where it and ZeroOneLoss.Score differ:

using Lodestar.Metrics;

bool[] truth = [true, false, true, false, true, true];
bool[] predicted = [true, false, false, true, true, true];

double loss = HammingLoss.Score(truth, predicted, labelCount: 3);  // => 0.3333…

Two of the six labels are wrong. ZeroOneLoss.Score reads 1 on the same input, because both rows carry a mistake.

Applies to — net10.0, netstandard2.0.

See also — ZeroOneLoss.Score, Accuracy.Score, the Python equivalence table.