Reproduce “FastSHAP: Real-Time Shapley Value Estimation” 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
Neil Jethani, Mukund Sudarshan, Ian Covert, Su-In Lee, Rajesh Ranganath
FastSHAP, developed by researchers at NYU and University of Washington, introduces a method for real-time Shapley value estimation by training an amortized explainer model. It achieves up to a 600x speedup over KernelSHAP while maintaining explanation accuracy, making principled model explanations practical for complex, high-dimensional black-box models.
Key insights
- Shapley values can be effectively amortized by a learned model trained via a surrogate objective, bypassing the need for computationally intractable direct ground-truth computation.
- The weighted least squares characterization of Shapley values provides a robust and optimizable training objective for a learned explainer model.
- Techniques like additive efficient normalization and gradient variance reduction (e.g., paired sampling) are crucial for training stability and ensuring accurate learned Shapley value estimates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.