Post-hoc eXplainable Artificial Intelligence Methods: Counterfactuals and XCBR Applications

Abstract

This thesis addresses the growing need for transparency and user-centric approaches in Explainable Artificial Intelligence (XAI), particularly in applications where AI makes high-stakes decisions. As the adoption of AI accelerates in critical fields, such as healthcare and finance, the need for explanations that are both reliable and understandable has amplified. Counterfactual explanations, which explore «what-if» scenarios, provide actionable insights that align well with human reasoning, making them highly effective for fostering trust and demonstrating causality in AI systems. This thesis is structured around three core research questions: (1) how integrating domain knowledge and human-centered design can enhance the relevance and trustworthiness of explanations, (2) the development of methods for generating stable, contextually meaningful counterfactuals, and (3) improving comprehensibility and user satisfaction in XAI systems. To address the challenges in these research questions, the thesis introduces several key contributions, including but not limited to PertCF, a perturbation-based method designed for generating counterfactual explanations that balance stability with minimal changes, and the CEval toolkit, a comprehensive evaluation framework developed to assess the fidelity, proximity, diversity, and robustness of instance-based explanations across various metrics. Empirical studies further validate these contributions, showing that counterfactuals can be tailored to a variety of data types, including tabular and image data while aligning with user preferences when explanations are adaptable to user needs. These findings underscore the value of user-centered XAI, establishing a foundation for future advancements in creating scalable, multimodal, and user-focused explainability methods. The approaches developed in this thesis offer insights for researchers and practitioners working toward transparent and trustworthy AI systems.