Beyond Shapley: Efficient Computation of Asymmetric Shapley Values

Short summary

Asymmetric Shapley Values (ASV) overcome computational barriers in model explainability by incorporating causal knowledge into feature attribution. The paper proves ASV can be computed in polynomial time where standard SHAP is #P-hard, introducing practical algorithms for tree-structured and arbitrary DAG causal graphs. Experimental validation on realistic structures demonstrates viability of previously intractable explainability methods.

• •ASV incorporates causal knowledge into feature attribution, addressing limitations of standard Shapley values
• •Polynomial-time computation achievable in cases where SHAP is computationally intractable
• •Practical algorithms provided for both tree-structured and general DAG causal graphs

Generated with AI, which can make mistakes.