Efficient sampling algorithms for plurality-voting power indices

Develop efficient sampling algorithms for estimating the Banzhaf and Shapley power indices in plurality-voting ensembles with random tie-breaking, especially for the multiclass classification setting, potentially by reusing samples across different base classifiers.

Background

The paper introduces exact algorithms for computing Banzhaf and Shapley values in plurality-voting ensembles with random tie-breaking. Although these algorithms are efficient in the binary case and polynomial-time in the multiclass case, the authors also implement simple Monte Carlo baselines. Those baselines require many samples and can provide inaccurate estimates, particularly when the power-index values are small.

The unresolved direction concerns designing more efficient and sample-efficient randomized estimators, with particular emphasis on multiclass ensembles. Reusing samples across different base classifiers is identified as one possible way to reduce computational cost and improve estimation quality. The same issue is revisited in the additional experimental results, where the authors again describe efficient sampling for the multiclass formulation as an open direction.

References

We remark that obtaining efficient sampling algorithms for our formulation remains an interesting and challenging open direction, especially for the multiclass case.

— Exact power indices for plurality-voting ensembles  (2610.17980 - Sarpe et al., 13 Oct 2026) in Appendix, Section appsec:bl, “Monte-Carlo sampling baseline”