This thesis explores a new approach to approximating decision problems in abstract argumentation using Graph Convolutional Networks (GCN). It demonstrates that such an approach can reach well-balanced accuracy levels above 90 \% across a range of different decision problems, argumentation semantics, and benchmarks.
This thesis develops a new Deep Neural Network (DNN) architecture adapted from the classic GCN that better addresses the specific issues found in abstract argumentation. Likewise, it develops a training approach that produces superior results for abstract argumentation data sets by introducing structured randomness and dynamic adaptation to the training data.
Then, the thesis systematically applies this architecture to a large argumentation dataset across the main argumentation semantics used in the biannual ICCMA competition. It evaluates the performance of the model in a variety of different settings and across benchmarks, size bands, and model variants. The main models show good performance in the majority of cases, although there is some variation.
Having created the core model, the thesis goes on to explore additional extensions of the core work. This first focuses on combining the approximate approach with exact approaches using a deterministic algorithm and a SAT solver, showing an improvement by solving six additional hard instances relative to existing solvers.
Second, we explore a visualisation approach that can give new insights into argumentation graphs by applying a dimensionality reduction technique to weights from the trained GCN models, showing new insights in explaining benchmark performance.
Finally, we explore using the same basic architecture to address another problem that can be structured using abstract argumentation. In this case, we apply the approach to the prediction of misinformation in tweets and achieve good performance on a key dataset.
