A1771
Title: Seeing through correlations: Disentangled feature importance
Authors: Jin-Hong Du - The University of Hong Kong (Hong Kong) [presenting]
Kathryn Roeder - Carnegie Mellon University (United States)
Larry Wasserman - Carnegie Mellon University (United States)
Abstract: Quantifying Feature Importance with valid statistical uncertainty is central to interpretable machine learning, yet classical model-agnostic methods often fail under Feature correlation, producing unreliable attributions and compromising statistical inference. Existing approaches, such as Shapley values and leave-one-covariate-out, are designed for Feature selection and vulnerable to correlation distortion, limiting their robustness on model interpretation. Disentangled Feature Importance (DFI), a model-agnostic framework, resolves these limitations by combining principled statistical inference with computational flexibility. DFI leverages entropic optimal transport to learn flexible disentanglement maps and provide an interpretable pathway for understanding how Importance is attributed through the data's correlation structure. The framework generalizes to flow matching and differentiable loss functions, enabling statistically valid Importance assessment for black-box predictors in both regression and classification. Statistical inference theory enables valid confidence intervals and hypothesis testing with Type I error control. Empirical results on synthetic and biomedical datasets show that DFI delivers substantially higher statistical power than removal-based and conditional permutation methods, while maintaining robust, interpretable attributions under severe Feature interdependence.
