Distribution Shift Detection

The shift module provides methods for detecting and quantifying distribution shifts between training and deployment data.

Base Class

BaseShiftDetector()

Abstract base class for distribution shift detection methods.

Shift Detectors

MMDShiftDetector([sigma])	Kernel Maximum Mean Discrepancy with Gaussian (RBF) kernel.
EnergyShiftDetector()	Energy distance – Szekely & Rizzo, 2013.
KSShiftDetector()	One-dimensional Kolmogorov–Smirnov test (per feature, max statistic).
ClassifierShiftDetector([clf_factory, device])	Train a logistic regression to separate reference and test.
BBSDDetector	alias of ClassifierShiftDetector
LabelShiftDetector(num_classes[, calibrated])	Black-Box Shift Detection for label shift.
ImportanceWeightingShift([method, alpha])	Importance weighting for covariate shift adaptation.

Metrics

energy_distance(x, y)	Szekely–Rizzo energy distance, O(n²) naive implementation.
total_variation(p, q[, eps])	Total variation between discrete distributions p and q.
population_stability_index(p_hist, q_hist[, eps])	Classic tabular PSI used in credit scoring.
wasserstein_distance(x, y[, p, max_iter, ...])	Wasserstein distance (Earth Mover's Distance) between two empirical distributions.
sliced_wasserstein_distance(x, y[, ...])	Sliced Wasserstein distance between two empirical distributions.

Visualization

plot_feature_histograms(ref, test[, ...])	Overlay 1-D histograms for a handful of features.
plot_embedding_space(ref_emb, test_emb[, ...])	Visualize reference and test embeddings in 2D space.
plot_confidence_distributions(...[, bins, show])	Compare model confidence distributions between reference and test data.
plot_shift_severity(severity_values, ...[, ...])	Plot shift scores against a severity measure (e.g., rotation degrees, time).
plot_ks_statistics(ks_stats[, ...])	Bar chart of per-feature KS statistics showing which features shifted most.