Metric helpers

The tsseg.metrics namespace provides classes implementing common segmentation quality measures. Metrics share a __call__ signature accepting predictions and ground-truth annotations, and many expose extra helpers for per-segment analysis.

Change-point metrics

• tsseg.metrics.F1Score – balanced F1 score computed from matched change points with tolerance windows.

• tsseg.metrics.Covering – overlap-based covering score between true and predicted segments.

• tsseg.metrics.HausdorffDistance – worst-case deviation between true and predicted change points.

• tsseg.metrics.GaussianF1 – F1 score using Gaussian-weighted tolerance around true change points.

• tsseg.metrics.BidirectionalCovering – symmetric segment-overlap score measuring both precision and recall coverage.

State-labelling metrics

• tsseg.metrics.AdjustedRandIndex – adjusts the Rand index for chance agreement.

• tsseg.metrics.AdjustedMutualInformation – normalized mutual information corrected for chance.

• tsseg.metrics.NormalizedMutualInformation – mutual information normalized by joint entropy.

• tsseg.metrics.WeightedAdjustedRandIndex – class-imbalance aware variant of the adjusted Rand index.

• tsseg.metrics.WeightedNormalizedMutualInformation – weighted version of the normalized mutual information.

• tsseg.metrics.StateMatchingScore – Hungarian matching-based score that first aligns predicted and true states before computing accuracy.

Base protocol

All metrics inherit from tsseg.metrics.BaseMetric, which defines the common update/compute contract used by collection utilities. Refer to Getting started for integration examples.