Reproduce “Explaining Models by Propagating Shapley Values of Local Components” 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
Hugh Chen, Scott Lundberg, Su-In Lee
Researchers developed an extended DeepSHAP framework that efficiently approximates Shapley values for diverse model architectures, including mixed model types and loss functions, by theoretically grounding the use of background distributions and improving propagation rules. This method provides unbiased and contextually relevant feature attributions, demonstrating effectiveness in medical prediction tasks and for model stacks.
Key insights
- DeepLIFT's propagation rules (Rescale and RevealCancel) can be explicitly reinterpreted and aligned with SHAP value calculations, enabling efficient layer-wise attribution.
- Averaging single-reference SHAP values over a background distribution provides a theoretically sound approach to obtaining true SHAP values for that distribution, mitigating bias from single baselines.
- Customizing background distributions allows for generating contextually relevant explanations, which is particularly valuable for domain experts (e.g., in medical applications).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.