Metrics meanabsoluteerror peroutput

MeanAbsoluteError.PerOutput

One mean absolute 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 — three samples, two outputs, the second predicted twice as badly as the first.

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 = MeanAbsoluteError.PerOutput(yTrue, yPred, outputCount: 2);
double first = perOutput[0];    // => 0.5
double second = perOutput[1];   // => 1

Remarks — scikit-learn's multioutput="raw_values". Reach for it whenever the outputs are not interchangeable — a model predicting both a price and a delay has no useful average of the two, and Score's plain mean of them is a number in no units at all.

outputWeights on Score is the middle ground when the outputs are commensurable but not equally important, and it is applied to exactly this array. There is no separate weighted-array form because weighting an array you are not reducing means nothing.

The trap is outputCount silently succeeding. It is only checked against the total length, so passing 2 for data that is really three outputs wide will slice the span into pairs and return two numbers computed from the wrong columns.

Applies to — net10.0, netstandard2.0.

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