v1.6.0 (2026-07-06)

Highlights of new Features

• adds the ShaplEIG approximator in shapiq.approximator.shapleig for Shapley value estimation via Bayesian experimental design (Rundel et al., 2026) #548
• adds the OddSHAP approximator in shapiq.approximator.regression for fast first-order Shapley value estimation via odd-only Fourier regression (Fumagalli et al., 2026) #522

Introducing ShaplEIG #548, ICML'26

Adds ShaplEIG, a Bayesian experimental design approximator for Shapley values (Rundel et al., 2026, "ShaplEIG: Bayesian Experimental Design for Shapley Value Estimation"). A Gaussian process surrogate with a weighted Hamming product kernel is fit on the queried coalition values, and the next coalition to evaluate is selected by maximizing the closed-form expected information gain (EIG) about the Shapley values. The Shapley structure is handled through elementary-symmetric-polynomial identities, so the full 2^n coalition space is never enumerated and each iteration costs only polynomial time in n. The returned InteractionValues hold the posterior-mean Shapley value estimates; approximate_with_variance additionally returns their marginal posterior variances. It is exported as shapiq.ShaplEIG.

ShaplEIG requires the new optional shapleig dependency group (pip install shapiq[shapleig] — torch, gpytorch, botorch, linear_operator); these optional dependencies are imported lazily.

Introducing OddSHAP #522, ICML'26

Adds OddSHAP, a first-order Shapley value estimator that exploits the odd structure of the Shapley value (Fumagalli et al., 2026, "An Odd Estimator for Shapley Values"). It combines paired sampling, sparse odd-interaction screening via a surrogate tree model whose exact Fourier (Walsh) coefficients are extracted, and a constrained odd Fourier regression that enforces the efficiency axiom exactly. The surrogate defaults to LightGBM (the paper's configuration) and transparently falls back to a scikit-learn DecisionTreeRegressor when LightGBM is unavailable. It is exported as shapiq.OddSHAP and restricted to index="SV", max_order=1.

shapiq Benchmark #521

Adds the shapiq_benchmark package, a framework to configure and run approximation benchmarks from either string tags or objects. It ships benchmark types for local XAI, image, TabPFN, path-dependent, and interventional games, a set of approximation-quality metrics, and Optuna-based hyperparameter optimization scripts with pre-computed best-parameter configs.

The package installs next to shapiq (like shapiq_games) and imports on a plain pip install shapiq. Its model backends are optional and imported lazily; install them with pip install shapiq[benchmark] (optuna, tabpfn, lightgbm, xgboost) to build the corresponding models or run the hyperparameter optimization.

Documentation

• includes the nearest-neighbor explainers (KNNExplainer, WeightedKNNExplainer, ThresholdNNExplainer) in the API documentation and the shapiq.explainer package exports #558

New Contributors

• @david-rundel made their first contribution in #548
• @Sara-ne made their first contribution in #522

Full Changelog: v1.5.2...v1.6.0