Enhancing explainability in real-world scenarios: Towards a robust stability measure for local interpretability

Eduardo Sepúlveda, Félix Vandervorst, Bart Baesens, Tim Verdonck

Abstract: Highlights•Measuring stability to enhance model reliability in Explainable AI (XAI).•Highlighting model trust in finance, e.g., fraud detection.•Proposing a new stability metric for local interpretability in anomaly detection.•Assessing feature ranking consistency under slight data changes.•Showing how stability improves performance and interpretability in tests.