Beyond SHAP - Functional Decomposition as a Common Language for Explainable AI in Medicine
Abstract:
Explainability has become a prerequisite for machine learning in medicine, partly for regulatory reasons, but above all because insight into why a model predicts what it predicts is what generates biomedical knowledge. In practice, researchers pick from a toolbox of methods such as feature importance, partial dependence and SHAP, whose relationships are poorly understood and which can disagree substantially on the same model and data. This talk presents functional decomposition as a common language for these methods. Writing a prediction function as a sum of main effects and interaction effects of increasing order, and fixing a suitable identification constraint, yields partial dependence plots and SHAP values from one and the same object, which unifies global and local explanations. This clarifies which question each method actually answers, why SHAP values can mislead when interactions are present, and how interactions can be quantified explicitly instead of being absorbed into single-feature attributions. I also present fast tree-based algorithms that estimate arbitrary components of the decomposition, and hence several explanations at once. I illustrate the approach with examples from large population-based cohort studies, and then turn to time-to-event outcomes, where the non-additivity of hazard and survival functions invalidates standard additive explanations. Separating time-dependent from time-independent effects reveals how much a factor contributes and also when, including in multimodal settings combining imaging and clinical data.
About Marvin:
Marvin N. Wright builds machine learning that doesn’t just predict, but explains. His research sits where modern machine learning meets classical statistics, and centers on some of the recurring challenges in health data science: how to make complex models interpretable, how to generate realistic synthetic data, and how to draw valid statistical conclusions from algorithms designed only to predict. He is the creator of ranger, a widely used random forest implementation, and his work on explainable AI, survival analysis, and generative modeling has appeared in venues including ICML, AISTATS, Statistics in Medicine, and Bioinformatics. A common thread runs through it: pushing machine learning in the health sciences beyond prediction, toward transparency, trust, and valid inference. Marvin is Professor of Machine Learning in Statistics at the University of Bremen and heads the Department of Statistical Methods in Epidemiology at the Leibniz Institute for Prevention Research and Epidemiology – BIPS, Germany, where he leads an Emmy Noether Research Group funded by the German Research Foundation (DFG).