New algorithm computes causal Shapley values in polynomial time
Asymmetric Shapley Values promise faster, more accurate AI explanations on causal graphs.
Explainability in machine learning often relies on Shapley values, but computing them exactly is #P-hard in many contexts. A new paper by Ezequiel Companeetz, Santiago Cifuentes, and Sergio Abriola tackles this via Asymmetric Shapley Values (ASV), which incorporate causal knowledge from a graph. The authors prove that for rooted directed trees, ASV can be computed exactly in polynomial time—a stark improvement over standard SHAP. They introduce equivalence classes over topological orderings to reduce computation further.
For general causal DAGs, the team develops an approximation algorithm that samples topological orderings uniformly at random. They leverage existing sampling mechanisms and simpler alternatives. The 18-page paper includes 6 figures and experimental results on realistic causal structures, confirming practical viability. This work bridges causality and interpretability, enabling faster, more faithful explanations for AI systems where causal dependencies matter.
- Exact ASV computation is polynomial-time on rooted directed trees, vs. #P-hard for classic SHAP.
- Equivalence classes over topological orderings reduce computation time for ASV.
- Approximation algorithm for arbitrary causal DAGs uses uniform sampling of topological orderings.
Why It Matters
Faster causal explainability for high-stakes AI decisions, from healthcare to finance.