The Smoothing Pathology: Structural Bias in Conditional Shapley Values Under Gaussian Imputation

This work studies a structural bias affecting conditional Shapley values when conditional expectations are estimated through Gaussian generative imputation models. We show that, under multimodal conditional distributions, models trained with an L2 reconstruction loss and a unimodal Gaussian decoder tend to converge toward the conditional mean. This produces samples located in low-density regions, a phenomenon we call the smoothing pathology. We analyze how this imputation bias propagates to conditional Shapley values and may lead to misleading feature attributions. The study is motivated by VAEAC-based conditional imputation for explainable AI and includes theoretical analysis, synthetic examples, and connections with conditional SHAP estimation