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

Machine learning (ML) models are powerful tools for detecting complex patterns, yet their 'black-box' nature limits their interpretability, hindering their use in critical domains like healthcare and finance. Interpretable ML methods aim to explain how features influence model predictions but often focus on univariate feature importance, overlooking complex feature interactions. Although recent efforts extend interpretability to feature interactions, existing approaches struggle with robustness and error control, especially under data perturbations. In this study, we introduce Diamond, a method for trustworthy feature interaction discovery. Diamond uniquely integrates the model-X knockoffs framework to control the false discovery rate, ensuring a low proportion of falsely detected interactions. Diamond includes a non-additivity distillation procedure that refines existing interaction importance measures to isolate non-additive interaction effects and preserve false discovery rate control. This approach addresses the limitations of off-the-shelf interaction measures, which, when used naively, can lead to inaccurate discoveries. Diamond's applicability spans a broad class of ML models, including deep neural networks, transformers, tree-based models and factorization-based models. Empirical evaluations on both simulated and real datasets across various biomedical studies demonstrate its utility in enabling reliable data-driven scientific discoveries. Diamond represents a significant step forward in leveraging ML for scientific innovation and hypothesis generation.