Beyond Divination: Stabilizing the Interpretability of Machine Learning Algorithms

Machine learning models achieve increasingly strong predictive performance, yet their decision processes often remain opaque and difficult to understand. This lack of transparency limits their deployment in sensitive and high-stakes settings, where predictions must be explained to be trusted and used. Explainability encompasses approaches that aim to make model behavior more understandable to users. In particular, local feature attribution methods quantify the contribution of each input variable to a model’s output by assigning importance values, which constitute the explanation for a given prediction. Approaches based on the Shapley value are widely adopted within this family due to their strong theoretical guarantees. In practice, however, exact Shapley values are generally not tractable and must be estimated. Most existing approaches rely on stochastic sampling procedures, which introduce variability into the resulting explanations and thus compromise stability, leading identical inputs to receive different feature importance values and undermining the reliability of the results. To address this issue, this thesis introduces ST-SHAP, a feature attribution method designed to improve stability by limiting the effect of randomness in the estimation process, and StratoSHAP, which defines a family of deterministic feature attribution methods that remove randomness and yield fully stable results. Together, these contributions lead to more stable feature attributions and enable more consistent analysis of machine learning model outputs.