Core Concepts
Relation Attribution Explanations (RAEs) quantify the contribution of attacks and supports in Quantitative Bipolar Argumentation Frameworks (QBAFs) to explain the strength of arguments.
Abstract
The paper proposes a novel theory of Relation Attribution Explanations (RAEs) to offer fine-grained insights into the role of attacks and supports in quantitative bipolar argumentation. RAEs are based on Shapley values from game theory and aim to explain the strength of a topic argument by quantifying the contribution of each edge (attack or support relation) in the QBAF.
The key highlights and insights are:
RAEs satisfy several desirable properties adapted from Shapley value properties, such as Efficiency, Dummy, Symmetry, and Dominance. They also introduce new argumentative properties like Sign Correctness, Counterfactuality, Qualitative Invariability, and Quantitative Variability.
The satisfaction and violation of these properties theoretically show that RAEs provide reasonable and faithful explanations, which is crucial for explanation methods.
The authors propose a probabilistic algorithm to efficiently approximate RAEs, prove theoretical convergence guarantees, and demonstrate experimentally that it converges quickly.
Two case studies are carried out to evaluate and demonstrate the practical usefulness of RAEs in fraud detection and large language model explanation tasks. The case studies illustrate how RAEs can provide more fine-grained insights compared to argument-based attribution explanations.
Overall, the paper introduces a novel and principled approach to explain the strength of arguments in QBAFs by considering the contributions of individual attacks and supports.
Stats
The base scores of all arguments in the QBAF example are set to 0.5.
Under the DF-QuAD gradual semantics, the strength of the topic argument α is 0.8046875.
Quotes
"Relation Attribution Explanations (RAEs) look at every subset of edges (S ⊆ R) and compute the marginal contribution of r (σS∪{r}(α) - σS(α))."
"RAEs satisfy several desirable properties adapted from Shapley value properties, such as Efficiency, Dummy, Symmetry, and Dominance."
"The satisfaction and violation of these properties theoretically show that RAEs provide reasonable and faithful explanations, which is crucial for explanation methods."