Speaker
Description
Understanding the relative importance of genetic, molecular and environmental factors is crucial for interpretable prediction models in biomedicine and for targeted prevention. While classical regression-based approaches provide direct interpretability through model coefficients, flexible machine learning (ML) approaches such as random forests and neural networks typically rely on post-hoc importance measures. Many state-of-the-art interpretability tools, including Shapley values, Banzhaf values and Beta Shapley values, can be viewed as probabilistic measures of variable importance motivated by coalition game theory, yet their conceptual links to Bayesian modelling have remained underexplored.
In this work, we develop a unifying Bayesian variable selection perspective on probabilistic variable importance measures in ML. We show that classical importance measures arise naturally under different priors on the model (coalition) space that encode different preferences over model complexity: a uniform prior on the model space leads to Banzhaf values; a uniform prior on the model size corresponds to Shapley values; and hierarchical Beta-Binomial priors give rise to Beta Shapley values. The Bayesian viewpoint clarifies the assumptions, trade-offs and interpretability properties underlying each measure. Furthermore, we introduce novel probabilistic importance measures including empirical Bayes formulations. Finally, we illustrate how the Bayesian variable selection perspective on interpretable ML can facilitate computations via Markov Chain Monte Carlo (MCMC) approaches for approximating probabilistic importance measures in high-dimensional biomedical data applications.
85717603055