Member Junction
    Preparing search index...

    The evaluation record a model carries in MLModel.Lineage.evaluation: building it at training, reading it back.

    Index

    Constructors

    Properties

    DefaultMajorityBaselineRule: MajorityBaselineRule = 'enforce'

    The rule a pipeline gets when its ValidationStrategy does not name one (D2-C5d).

    Key: "evaluation" = 'evaluation'

    The Lineage key the record is stored under.

    MajorityBaselineRules: readonly MajorityBaselineRule[] = ...

    Every value MajorityBaselineRule accepts.

    NotEvaluatedAsForecast: "not evaluated as a forecast" = 'not evaluated as a forecast'

    The stored reason for a time-indexed model that was not tested on later periods.

    Methods

    • The record training writes for a pipeline's validation strategy. ValidationOutOfSample is required: it is the flag that lets validation Metrics count toward trust, so it never defaults to true. Pass MajorityBaselineRule when the caller has already parsed it with ParseMajorityBaselineRule; otherwise it is parsed here, which throws InvalidMajorityBaselineRuleError for an unknown value. A declared time index withholds trust (NotEvaluatedAsForecast) unless HoldoutRespectsTime says the model was tested on later periods than it trained on (a date-carved or rolling-origin holdout).

      Parameters

      • validation: ValidationStrategy
      • options: {
            HoldoutRespectsTime?: boolean;
            ImportanceKind?:
                | "none"
                | "coefficient"
                | "split_or_impurity"
                | "held_out_permutation"
                | null;
            MajorityBaselineRule?: MajorityBaselineRule;
            ValidationOutOfSample: boolean;
        }

      Returns EvaluationRecord