Reproduce “Interpretable & Explorable Approximations of Black Box Models” 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

Himabindu Lakkaraju, Ece Kamar, Rich Caruana, Jure Leskovec

BETA (Black Box Explanations through Transparent Approximations) describes complex classifiers with compact, globally applicable two-level decision sets. On a depression-diagnosis dataset, it achieved roughly 85% agreement with a five-layer neural network at lower explanation complexity than comparison methods, while user studies showed higher accuracy and faster interpretation, reaching 98.3% accuracy with feature-focused interactive explanations.

Key insights

  • Global explanations can be made more useful by allowing different decision logic in different neighborhoods rather than forcing one uniform rule system.
  • Reducing overlap between rules helps users identify a single rationale for an instance, while coverage and tie-breaking rules preserve operational completeness.
  • The formal optimization procedure has an approximately one-fifth guarantee for its objective, but this does not directly guarantee predictive accuracy, fairness, or human comprehension.

Summary by alphaXiv.

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