Error-controlled non-additive interaction discovery in machine learning models.

Area of Science:

• Artificial Intelligence
• Computational Biology
• Data Science

Background:

• Machine learning (ML) models excel at pattern detection but often lack interpretability due to their 'black-box' nature.
• Existing interpretable ML methods primarily focus on univariate feature importance, neglecting complex feature interactions.
• Current approaches for feature interaction interpretability lack robustness and effective error control, especially with data perturbations.

Purpose of the Study:

• Introduce Diamond, a novel method for trustworthy feature interaction discovery in machine learning.
• Address the limitations of existing methods in controlling false discoveries and handling non-additive interaction effects.
• Enhance the reliability of ML-driven scientific discovery and hypothesis generation.

Main Methods:

• Integrate the model-X knockoffs framework for rigorous false discovery rate (FDR) control.
• Employ a non-additivity distillation procedure to refine interaction importance measures and isolate non-additive effects.
• Ensure FDR control is preserved throughout the interaction discovery process.

Main Results:

• Diamond demonstrates robust feature interaction discovery across diverse ML models, including deep neural networks and transformers.
• Empirical evaluations on simulated and real biomedical datasets confirm Diamond's utility in enabling reliable data-driven discoveries.
• The method effectively isolates non-additive interaction effects, overcoming limitations of naive interaction importance measures.

Conclusions:

• Diamond significantly advances trustworthy feature interaction discovery in machine learning.
• The method facilitates reliable scientific innovation and hypothesis generation by providing interpretable and robust insights.
• Diamond enhances the applicability of ML in critical domains like healthcare and finance through improved interpretability.