From Log-Odds to Shapley Values: An Explanatory Geometry for the Weighted Naive Bayes Classifier

Title: From Log-Odds to Shapley Values: An Explanatory Geometry for the Weighted Naive Bayes Classifier Source: arxiv Reformulating weighted Naive Bayes via conditional log-odds induces a supervised distance equivalent to the L1 distance between analytical Shapley value vectors. This mapping links the model's geometric representation directly to local feature importance. Empirical testing using k-nearest neighbors validates these distances for predictive modeling and explanation consistency.