Reproduce “The Shapley Taylor Interaction Index” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kedar Dhamdhere, Ashish Agarwal, Mukund Sundararajan
Researchers from Google LLC propose the Shapley-Taylor Interaction Index, a novel method for attributing machine learning model predictions to interactions between subsets of features. The index generalizes the classic Shapley value, uniquely satisfies five desirable axioms including efficiency, and demonstrates a deep connection to Taylor series expansions, providing robust, intuitive per-instance explanations of complex model behavior in tasks like sentiment analysis and question answering.
Key insights
- The Shapley-Taylor index is uniquely characterized by five axioms, including a crucial efficiency axiom ensuring attributions sum coherently to the total prediction.
- The index generalizes the classic Shapley value and is robust against the pitfalls (like amplified or incorrectly signed interactions) of prior Shapley interaction indices.
- A deep mathematical connection exists between the Shapley-Taylor index and the Taylor series expansion of a model's multilinear extension, providing a continuous interpretation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.