fiXAIt: A novel feature importance-based XAI tool for enhanced explainability, self-consistency, and computational efficiency?

Abstract

This paper proposes fiXAIt, a post-hoc, model-agnostic framework for feature-importance explanation, designed with practical applications in mind, including healthcare, finance and risk management. The main idea is to separate effects that come from the presence of a feature from those that come from its actual value. For this purpose, we define two measures, Feature Existence Impact (FEI) and Feature Value Impact (FVI), and we add an internal Self-Consistency (SC) metric so that the explanations stay in line with the model's behavior. In our experiments, SC was strongly associated with model stability (SC vs. CV-F1 Std r = -0.63). To ensure scalability, fiXAIt combines CBSFSA for targeted feature selection and ECFC for pruning redundant feature combinations. The framework was evaluated on five heterogeneous datasets and various classifier families, and models retrained on features selected by fiXAIt yielded test accuracies close to the corresponding full models. This shows that fiXAIt's feature selection fidelity is aligned with model performance. In addition, fiXAIt performs significantly faster than SHAP. In tests, fiXAIt decreased the average explanation time from roughly 33.6 s to 18.6 s, which is about a 45 % improvement. These gains make it much easier to perform repeated analyses or to work with large-scale datasets. By delivering dual-perspective attribution, an internal reliability indicator and lower computational cost, fiXAIt provides a practical and reproducible XAI solution for risk-sensitive applications.