js/ml/metrics/shared

type Label = string | number | boolean

A class label. Metrics compare labels by value, so any primitive works.

class MetricError extends Error {

Error thrown when metric inputs are malformed.

Covers length mismatches, empty inputs, non-finite values, probabilities outside [0, 1], unknown labels, and averaging modes that cannot be resolved from the data.

Constructors

class CompensatedSum {

Neumaier compensated summation.

Streaming metrics accumulate over arbitrarily many batches, where naive addition drifts once the running total dwarfs individual terms.

Methods

Getters

function requireSameLength( a: ArrayLike<unknown>, b: ArrayLike<unknown>, aName: string, bName: string, ): void

function requireNonEmpty(values: ArrayLike<unknown>, name: string): void

function requireFinite(values: ArrayLike<number>, name: string): void

function requireProbabilities(values: ArrayLike<number>, name: string): void

function requirePositiveInteger(value: number, name: string): void

function compareLabels(a: Label, b: Label): number

Total ordering over labels: numbers numerically, booleans false-first, strings lexicographically, and mixed types grouped by type name so the ordering stays stable regardless of insertion order.

function sortedLabelUnion(...sources: ReadonlyArray<ArrayLike<Label>>): Label[]

function resolvePositiveLabel(labels: readonly Label[], explicit?: Label): Label

Resolve which label counts as positive for a binary metric.

An explicit choice always wins. Otherwise {0, 1} resolves to 1 and {false, true} to true; anything else is ambiguous and must be stated.

function resolveScoreLabel(labels: readonly Label[], explicit?: Label): Label

Resolve the positive label for a probability- or score-based metric.

Unlike resolvePositiveLabel this does not require the positive class to actually appear: a batch can legitimately be all negatives, and refusing to score it would make streaming evaluation fail on unlucky shard boundaries.

function formatLabels(labels: readonly Label[]): string

function ratio(numerator: number, denominator: number): number

numerator / denominator, or 0 when the denominator is zero.

Undefined metric values are reported as zero throughout this module rather than as NaN, matching the convention established by scikit-learn.

function descendingOrder(scores: ArrayLike<number>): number[]

Ranks values best-first, breaking ties by original position so that equal scores keep input order and every derived curve is deterministic.