Realtime AI News
Beyond Shapley: Efficient Exact Computation of Asymmetric Shapley Values Achieved in New Work
A new arXiv paper presents a method for computing Asymmetric Shapley Values using causal graphs, achieving polynomial-time computation in scenarios where SHAP is #P-hard.
A paper titled 'Beyond Shapley: Efficient Computation of Asymmetric Shapley Values' has been published on arXiv. The research addresses the problem of explainability in machine learning models through feature attribution methods, proposing a variant called Asymmetric Shapley Values that incorporates causal knowledge into model-agnostic explanations via a causal graph.
The key contribution is showing that in certain contexts where the computation of SHAP is #P-hard, the exact computation of Asymmetric Shapley Values can be done in polynomial time.
The source is arXiv cs.AI (ID 2606.25103), published on June 25, 2026.
Why it matters
This work offers a breakthrough solution to the computational bottleneck in model explainability, making it possible to incorporate causal knowledge while maintaining efficient computation.