Metrics 0.2.0 meanabsolutepercentageerror 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.

MeanAbsolutePercentageError.PerOutput

One mean absolute percentage error per output, unreduced.

public static double[] PerOutput(ReadOnlySpan<double> yTrue, ReadOnlySpan<double> yPred, int outputCount = 1, ReadOnlySpan<double> sampleWeight = default)

Parameters — yTrue and yPred are the true and predicted values, row-major. outputCount is how many outputs each row holds, and sampleWeight weights the rows.

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.

Example — the same four numbers read as two samples of two outputs.

using Lodestar.Metrics;

double[] yTrue = [100.0, 50.0, 200.0, 25.0];
double[] yPred = [110.0, 45.0, 180.0, 30.0];

double[] perOutput = MeanAbsolutePercentageError.PerOutput(yTrue, yPred, outputCount: 2);
double first = perOutput[0];    // => 0.1
double second = perOutput[1];   // => 0.15…

Remarks — because this metric is already scale-free, its per-output array is one of the few here whose entries are directly comparable with one another: two outputs measured in different units still produce two percentages. That makes Score's plain mean of them meaningful in a way that MeanAbsoluteError.Score's is not.

The trap is that "scale-free" is a claim about the units, not about the data. An output whose truth sits near zero for some samples still explodes, and its entry will then dominate the mean over outputs exactly as it would dominate a mean over samples.

Applies to — net10.0, netstandard2.0.

See also — MeanAbsolutePercentageError.Score, MeanAbsoluteError.PerOutput, the Python equivalence table.