LightSGCN: Powering Signed Graph Convolution Network for Link Sign Prediction with Simplified Architecture Design

Abstract

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. In this work, we aim to simplify the architecture of signed GNNs to make it more concise and appropriate for link sign prediction. We propose a simplified signed graph convolution network model called LightSGCN. Specifically, LightSGCN utilizes linear propagation based on the balance theory, a widely adopted social theory. Then, the linear combination of hidden representations at each layer is used as the final representations. Moreover, we also propose a tailored prediction function. These finally yield a simple yet effective LightSGCN model, which is more interpretable, easier to implement, and more efficient to train. Experimental results on four real-world signed graphs demonstrate that such a linear method outperforms the state-of-the-art signed GNNs methods with significant improvement in the link sign prediction task and achieves more than 100X speedup over the most similar and simplest baseline.