Argumentation frameworks serve as tools for representing and reasoning over information in various domains, for example by adding explainability to AI models performing classification. This work concentrates on Quantitative Bipolar Argumentation Frameworks (QBAFs), which model both support and attack relations and assign numeric strengths to arguments. Existing explanation approaches typically address the outcome of a single topic argument, whereas contrastive explanations focus on the difference between two topic arguments. We first propose a general form of contrastive attribution functions (CAFs) and list essential properties such as symmetry, additivity and locality. Then we instantiate CAFs using three techniques:
- removal: iteratively delete arguments or relations and observe outcome changes;
- gradients: measure sensitivity via derivatives of argument strengths;
- Shapley values: allocate contributions from a cooperative game perspective.
For each instantiation we discuss which properties hold and the computational cost. Finally, we demonstrate the usefulness of contrastive explanations in two real‑world settings: comparing two possible disease labels in a medical diagnosis scenario to aid clinicians, and contrasting decisions across demographic groups to uncover bias. These experiments show that contrastive explanations provide finer‑grained insight and complement single‑point explanations.
Review