Reproduce “CXPlain: Causal Explanations for Model Interpretation under Uncertainty” 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

Patrick Schwab, Walter Karlen

CXPlain reframes feature attribution as a supervised learning problem: an auxiliary explainer learns causal importance scores from the performance change caused by masking each input feature of a black-box model. With bootstrap ensembles for confidence intervals, it produced more focused explanations than several baselines, matched or exceeded competing methods on image benchmarks, and reduced explanation-time computation by orders of magnitude compared with LIME and SHAP.

Key insights

  • The causal objective encourages attributions to concentrate on features whose removal actually harms the model’s prediction, rather than merely correlating with the output.
  • Training incurs an upfront cost, but amortizes it across rapid subsequent explanations; evaluation speed approaches that of gradient methods while applying to arbitrary black-box predictors.
  • Even five bootstrap models yielded uncertainty estimates correlated with attribution rank errors, with stronger correlations as ensemble size increased.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.