Argumentative explainable AI (XAI) has been advocated by several in recent years, with an increasing interest on explaining the reasoning outcomes of Argumentation Frameworks (AFs). While there is a considerable body of research on qualitatively explaining the reasoning outcomes of AFs with debates, disputes, or dialogues, explaining the quantitative reasoning outcomes of Quantitative Bipolar AFs (QBAFs) under gradual semantics has not received much attention, despite the widespread use of QBAFs in applications, such as voting polls and fraud detection, where explainability is crucial for ensuring comprehensibility and trust.
In this thesis, we contribute to filling this gap by proposing three novel theories of explanations. The first theory, Argument Attribution Explanations (AAEs), incorporates the spirit of feature attribution from machine learning in the context of QBAFs: whereas feature attributions identify the influence of features towards outputs of machine learning models, AAEs identify the influence of arguments towards a specific topic argument of interest. The second theory, Relation Attribution Explanations (RAEs), shifts the focus from arguments to the relations between arguments. RAEs measure the influence of relations towards a topic argument, providing a more fine-grained understanding than AAEs. The third theory is Counterfactual Explanations (CEs). Unlike AAEs and RAEs which explain the existing outcome, CEs suggest how to change the outcome to a desired one by modifying the QBAF in a cost-effective manner.
To evaluate these explanations, we theoretically study the desirable properties of the three proposed theories of explanations, including some new ones and some partially adapted from the literature to our setting. Additionally, we empirically validate the performance of these explanation methods, focusing on aspects such as scalability and robustness. Finally, we demonstrate the applicability of our proposed explanations by carrying out several case studies in various scenarios, such as fake news detection and movie recommendation.