Metrics d2tweedie score

D2Tweedie.Score

The fraction of Tweedie deviance explained — sklearn.metrics.d2_tweedie_score.

public static double Score(ReadOnlySpan<double> yTrue, ReadOnlySpan<double> yPred, double power = 0, ReadOnlySpan<double> sampleWeight = default, ZeroDivision zeroDivision = ZeroDivision.NaN)

Parameters — yTrue and yPred are the true and predicted values, of the same length. power selects which deviance is explained; TweedieDeviance has the table of regimes and what each admits, and this applies exactly the same rules. sampleWeight is one weight per sample, or empty — the default. zeroDivision decides the answer for fewer than two samples, the one case scikit-learn leaves undefined; the default reproduces its nan.

Returns — double. 1 for a perfect prediction, 0 for one no better than predicting the weighted average of the truth, and negative below that. Unbounded below.

Exceptions — ArgumentOutOfRangeException when power lies in (0, 1). ArgumentException when the lengths disagree, the input is empty or non-finite, or an operand is outside the regime's domain. UndefinedMetricException when every truth is the same value — the constant baseline is already perfect, so there is nothing to explain — or when there are fewer than two samples and zeroDivision is ZeroDivision.Throw.

Example — the worked case, read as a normal and then as a Poisson.

using Lodestar.Metrics;

double[] truth = [1.0, 2.0, 3.0, 4.0];
double[] predicted = [1.5, 2.5, 2.0, 4.5];

double normal = D2Tweedie.Score(truth, predicted);  // => 0.65

using Lodestar.Metrics;

double[] truth = [1.0, 2.0, 3.0, 4.0];
double[] predicted = [1.5, 2.5, 2.0, 4.5];

double poisson = D2Tweedie.Score(truth, predicted, power: 1.0);  // => 0.6302…

The first is R2.Score on the same input, to the last bit.

Remarks — scikit-learn warns and returns nan below two samples rather than refusing, which ZeroDivision.NaN reproduces; pass ZeroDivision.Zero, One or Throw for another answer, exactly as R2.Score takes it for the same case.

A constant truth is the other undefined case and is not governed by zeroDivision: it always throws, because the reference always raises there. D2AbsoluteError.Score answers 0 on that input instead, which is its own reference's behaviour.

Applies to — net10.0, netstandard2.0.

See also — TweedieDeviance.Score, R2.Score, D2AbsoluteError.Score, the Python equivalence table.