LightSGCN
Haoxin Liu · Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval · 2022
With both positive and negative links, signed graphs exist widely in the real world. Recently, signed graph neural networks (GNNs) have shown superior performance in the most common signed graph analysis task, i.e., link sign prediction. Existing signed GNNs follow the classic nonlinear-propagation paradigm in unsigned GNNs. However, several recent studies on unsigned GNNs have shown that such a paradigm increases training difficulty and even reduces performance in various unsigned graph analysis tasks. Meanwhile, most of the public real-world signed graph datasets do not provide node features. These motivate us to consider whether the existing complex model architecture is suitable.