Amélioration de l’interprétabilité des explications de SHAP grâce à la découverte de sous-groupes

The integration of predictive models in medicine requires understandable explanations to support clinical decision-making. This preliminary study introduces a post-hoc, model-agnostic approach that combines SHAP with DS to generate explicit IF–THEN rules. This combination provides both local and global explanations, while offering more precise insights than SHAP feature importance and a richer understanding of interactions between variables. Experiments conducted on four medical datasets demonstrate broad coverage and accurate class characterization, with high WRAcc and lift values. The associated local explanations achieve over 90% fidelity for binary models. Although developed in a medical context, the approach can be applied to any domain requiring intelligibility and trust in predictive models.

Conference paper