Effects of Graph Neural Network Aggregation Functions on Generalizability for Solving Abstract Argumentation Semantics

This paper investigates the effects of different graph neural network aggregation functions on the generalizability of these models for solving abstract argumentation semantics. We systematically compare the performance of Sum, Mean, Max, Attention, and Convolution aggregation functions on predicting sceptical argument acceptance under the preferred semantics. Our experiments utilize a variety of benchmark datasets to evaluate scalability and out-of-domain generalization. Results show that while most aggregators perform well on scalability, Max and Attention aggregators significantly outperform others on generalization to data not seen during training. This study provides valuable insights into the design of accurate and robust GNN-based approximate solvers for abstract argumentation frameworks, emphasizing the importance of the aggregation function.