Metrics 0.2.0 r2 peroutput

Lodestar.Metrics 0.2.0. 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.

R2.PerOutput

One score per output, unreduced.

public static double[] PerOutput(ReadOnlySpan<double> yTrue, ReadOnlySpan<double> yPred, int outputCount = 1, ReadOnlySpan<double> sampleWeight = default, bool forceFinite = true, ZeroDivision zeroDivision = ZeroDivision.NaN)

Parameters — yTrue and yPred are the true and predicted values, row-major. outputCount is how many outputs each row holds, sampleWeight weights the rows, forceFinite answers a truth of zero variance, and zeroDivision answers fewer than two samples.

Returns — a fresh double[] of outputCount entries, in column order.

Exceptions — ArgumentException when a length disagrees with the shape, the input is empty, or it holds a non-finite value; ArgumentOutOfRangeException when outputCount is below one; UndefinedMetricException when there are fewer than two samples and zeroDivision is ZeroDivision.Throw.

Example — three samples, two outputs.

using Lodestar.Metrics;

double[] yTrue = [0.5, 1.0, -1.0, 1.0, 7.0, -6.0];
double[] yPred = [0.0, 2.0, -1.0, 2.0, 8.0, -5.0];

double[] perOutput = R2.PerOutput(yTrue, yPred, outputCount: 2);
double first = perOutput[0];    // => 0.9654…
double second = perOutput[1];   // => 0.9081…

Remarks — because R2 is unitless, this is the multioutput array whose entries really are comparable with one another, which is what makes it the honest way to look at a model predicting several unrelated things. MeanSquaredError.PerOutput cannot do that.

There is one shape divergence from scikit-learn, and it is stated rather than hidden: on fewer than two samples with more than one output, this returns one NaN per output, where r2_score returns a single scalar nan before it ever consults multioutput. No number differs — every scalar-returning path here still gives NaN — and a one-element array would break this method's own contract of one value per output.

The trap is the same one Score has, one level down: a negative entry is not a bug. It means that output is predicted worse than its own mean would be, and on a multioutput model that is usually one column with almost no variance rather than a broken model.

Applies to — net10.0, netstandard2.0.

See also — R2.Score, R2.VarianceWeighted, ExplainedVariance.PerOutput, the Python equivalence table.