• Explainable Artificial Intelligence

Understanding the decision-making process of complex machine learning models remains a critical challenge, particularly for monolithic models that lack inherent interpretability. To address this issue, we propose an input-diversified ensemble approach based on counterfactual-derived feature importances. Our method leverages an ensemble of models trained on diverse subsets of the data to approximate the feature importance of the original monolithic model. By diversifying the training inputs, our ensemble captures a range of counterfactual explanations and then aggregates them to effectively approximate the feature importance of the original monolithic model. The core idea behind our approach is that different data subsets highlight different aspects of feature relevance, leading to a more robust approximation when aggregated. Each model within the ensemble is trained independently on a subset of the original data, varying their distribution and the balance between the classes, ensuring that the learned feature importances are not biased by the global distribution but rather capture local variations. This enables a more interpretable understanding of feature contributions without requiring access to the full complexity and training data of the monolithic model itself.