Op weg naar betekenisvolle post-hoc verklaringen voor eindgebruikers van voorspellende modellen

This thesis advances the understanding and practical use of local post-hoc explanations in machine learning. The first part investigates their interpretation: SHAP and LIME are compared to classical local linear approximations, showing that they capture broader aspects of a model's decision boundary and must therefore be interpreted differently. A new decomposition of SHAP values is also introduced, enabling a more complete treatment of feature dependencies. The second part explores how explanations can be made interactive, allowing end-users to correct explanations and thereby improve model performance and sample efficiency. Model-agnostic interactive implementations of SHAP and Occlusion are presented, followed by interactive adaptations of gradient-based methods for neural networks. Experiments across simulated and real-world datasets demonstrate that these interactive explanations can simultaneously improve both model predictions and explanation quality. Interestingly, simple methods often not only matched but even outperformed more sophisticated ones. Overall, this thesis highlights both conceptual and practical advances: it clarifies how local post-hoc explanations should be interpreted and shows how interactive frameworks can transform them from passive interpretability tools into active means of improving machine learning models.