Reproduce “A Confidence Machine for Sparse High-Order Interaction Model” on OpenResearch

About this paper

Diptesh Das, Eugene Ndiaye, Ichiro Takeuchi

Researchers from Nagoya University, Georgia Institute of Technology, and RIKEN developed "Homotopy Mining," an algorithm that computes exact full-Conformal Prediction (CP) sets for Sparse High-Order Interaction Models (SHIMs). This method provides statistically more compact confidence intervals and predictive accuracy comparable to complex black-box models, while making the computation tractable for these interpretable models.

Key insights

• Exact full-CP can be made computationally tractable for a class of complex, interpretable models like SHIMs by combining homotopy methods with efficient search and pruning strategies.
• Homotopy methods can be effectively extended beyond fixed feature sets, integrating with pattern mining techniques to handle the combinatorially large feature spaces inherent in high-order interaction models.
• SHIMs offer a valuable balance, providing predictive accuracy competitive with complex black-box models while maintaining interpretability, and when coupled with full-CP, delivering statistically superior and more precise uncertainty quantification.

Summary by alphaXiv.