SGNN: A New Method for Learning Representations on Signed Networks

Ji Jin, Jialin Bi, Cunquan Qu

Abstract: Graph Convolutional Neural Networks (GCNNs) have emerged as a powerful tool for processing graph-structured data and achieving outstanding performance in various scenarios, including node classification, link prediction, and graph visualization. However, in signed networks, which contain links with logically opposite relations, two types of associations can more accurately capture relationships in the real world. In this paper, we propose a novel optimization objective for learning representations in signed networks by capitalizing on the structural balance theory. Additionally, we introduce a new message propagation framework and develop a scalable signed graph neural network model, SGNN, which significantly outperforms existing signed graph embedding models when applied to the link classification task on four empirical networks. Moreover, our message propagation framework enables the construction of deeper GCNNs. Overall, our study presents a new approach for learning meaningful representations on signed networks, utilizing a novel message propagation framework and optimization objective that captures the complex relationships between nodes in a signed graph.