Contrastive Explanations Enhance AI Model Interpretability in Argumentation

Argumentation frameworks serve as valuable tools for representing and reasoning with information across various contexts, including their use in augmenting AI models for classification tasks, notably enhancing explainability. This research introduces contrastive explanations specifically for Quantitative Bipolar Argumentation Frameworks (QBAFs), a particular type of formalism. Unlike most existing QBAF explanations, which focus on clarifying the reasoning outcome of a single argument of interest, contrastive explanations aim to elucidate the differences between two distinct topic arguments. The paper outlines a general form for contrastive attribution functions (CAFs) and establishes a set of fundamental properties that these functions should satisfy. It then presents specific CAFs derived from methods such as removal, gradients, and Shapley-values, and thoroughly examines their properties. To illustrate the practical utility of contrastive explanations, the authors demonstrate their application in critical domains like healthcare decision-making and the identification of biases within AI systems.