Beyond feature attribution: Recovering latent interaction structures in machine learning models via localized CBFI

Overview & Synopsis

Background: Understanding individual predictions in complex tabular models remains challenging, as existing additive frameworks like SHAP and LIME often obscure structural logic by absorbing nonlinear interactions into individual feature scores. Exposing these latent interaction structures is essential for automated auditing in high-stakes environments where understanding the logic of model behavior is as essential as predictive accuracy. Methods: To address this limitation, we introduce the localized case-based feature importance (CBFI) framework. Unlike traditional numerical attribution, our method decomposes instance-specific predictions into structurally distinct components, effectively separating independent main effects (G1) from pure interactions (G4) to capture explicit feature synergies and provide a principled approach for diagnosing feature combinations. Results: Comprehensive experiments on wine quality classification and medical insurance regression demonstrate that localized CBFI successfully disentangles complex interactions hidden by additive models. Furthermore, fidelity tests and synthetic ground-truth benchmarks confirm that the identified interactions correspond precisely to the nonlinear structural dependency patterns embedded in complex architectures, including gradient boosting and deep neural networks, thereby successfully shifting the explainable AI paradigm from simple numerical attribution to the recovery of interaction-driven contributions.