From Comparing SHAP and LIME to Combining Them: A Faithfulness-Guided Fusion and Routing Framework with Actionable Counterfactuals for Healthcare and Financial Risk Prediction

Abstract

Comparative studies of explainable AI in high-stakes tabular domains keep reaching the same conclusion: use a hybrid explanation approach that employs several explanation methods as needed. They stop, however, at that recommendation. They do not specify how to integrate the methods, which method to trust on a given model or case, or whether combining even beats picking the best single method. We answer these questions on a measure-by-measure basis. Four attribution methods (SHAP, LIME, permutation importance, and impurity importance) and a counterfactual generator are applied to a trained model. Each method receives a faithfulness score between 0 and 1, where 0 is no better than randomly shuffling the features and 1 is as faithful as the optimal method. We then combine the methods in two ways. The first is global fusion: a single weighted average over the entire dataset, compared against an oracle that selects the best possible weights. The second is a per-instance router that, for each case, selects the method that best explains that case. The faithfulness test uses only the model and does not require labels for the router. We also produce actionable counterfactuals and judge them by validity, proximity, plausibility, and feasibility. We evaluate on three public datasets spanning two domains with two model families: stroke screening (healthcare), heart disease (healthcare), and German credit risk (finance), using random forest and gradient boosting. The results are consistent. The best method is hard to predict in advance—it depends on the dataset and even on the model family. Averaging the methods is not safe, because it can score below the best single method. Global fusion helps only a little, and only case by case rather than on average. The faithfulness gains (up to +0.22 on the credit model, where every method otherwise fails) come from per-instance routing, which turns an untrusted explanation into a borderline-trustworthy one. Restricting counterfactuals to actionable features raises feasibility to 100% while keeping validity at its full value, whereas unconstrained generation can leave feasibility as low as 4%. All datasets are public and obtained directly from their original repositories.