Reproduce “Rule Generation for Classification: Scalability, Interpretability, and Fairness” on OpenResearch

About this paper

Tabea E. Röber, Adia C. Lumadjeng, M. Hakan Akyüz, Ş. İlker Birbil

This research introduces "Rule Generation (RUG)," a scalable, LP-based optimization method for multi-class classification that generates concise, locally interpretable rule sets. The framework achieves competitive predictive performance compared to complex ensemble models while integrating generalized fairness constraints for diverse classification scenarios.

Key insights

• Leveraging decision trees with dual-variable-derived sample weights provides an effective and scalable heuristic for generating high-quality rules within a column generation framework.
• LP-based rule learning can natively support multi-class classification and simultaneously achieve high predictive accuracy, global interpretability (concise rules), and local interpretability (personalized explanations per sample).
• Fairness constraints, generalized for multi-class and multiple sensitive attributes, can be seamlessly integrated into the LP formulation, enabling the development of fair and transparent AI systems.

Summary by alphaXiv.