Reproduce “Asymmetric Shapley values: incorporating causal knowledge into model-agnostic explainability” on OpenResearch

About this paper

Christopher Frye, Colin Rowat, Ilya Feige

Researchers at Faculty and the University of Birmingham developed Asymmetric Shapley values (ASVs), a model-agnostic explainability framework that integrates causal knowledge by relaxing the traditional Symmetry Axiom. This extension provides more accurate and interpretable explanations, enabling capabilities like detecting unresolved discrimination in synthetic data, focusing attribution on initial steps in time-series data, and allowing for efficient, verifiable feature selection.

Key insights

• Incorporating even partial causal knowledge into explainability methods yields richer and more accurate insights into model behavior than purely correlational approaches.
• Asymmetric Shapley values bridge the gap between purely correlational explanations and full causal inference by allowing the integration of any known causal relationships.
• Relaxing the Symmetry Axiom is crucial for enabling a principled framework that can differentiate between the contributions of causally related features.

Summary by alphaXiv.