TreeSHAP-Importance-Weighted Feature-Group Fusion for Tabular Regression: Centered Additive Decomposition, Rade-Macher Complexity Control, and Attribution Stability

Interpretable feature-group fusion for tabular regression remains challenging because strong predictive models often lack explicit group-level attribution, complexity control, and attribution-stability guarantees. This paper proposes SHAP-Weighted Feature Fusion with Residual Mixing (SWFF-R), a closed-form convex fusion framework that combines a full-feature predictor with a SHAP-weighted blend of group-specific expert models. Group weights are obtained from temperature-controlled softmax transformations of group-level TreeSHAP importances, and the residual mixing coefficient is selected on a validation set to preserve predictive robustness. To avoid terminological confusion, we stress that these weights are derived from aggregated TreeSHAP importance scores of the full-feature model, not from group-level Shapley values recomputed by treating each feature group as a single player; the construction is therefore best described as TreeSHAP-importance-weighted feature-group fusion. Under fixed fusion weights and explicit centering, the proposed attribution satisfies Shapley-consistency properties for the induced centered additive group decomposition. We derive a Rademacher complexity upper bound for SWFF-R, provide a complementary minimax lower-bound calculation on a simplified linear subclass, and establish guarantees for temperature search and fixed-coefficient Lipschitz stability. Experiments on seven real-world tabular regression datasets and a separate synthetic 500K scalability stress test show that SWFF-R preserves predictive performance, yields point-estimate RMSE improvements on several datasets, and provides stable group-level attribution. Overall, SWFF-R offers a theoretically grounded framework for interpretable feature-group fusion in tabular regression.