Effective and lightweight representation learning for signed bipartite graphs

How can we effectively learn node representations in signed bipartite graphs? A signed bipartite graph consists of two node sets, where nodes of different types are positively or negatively connected. It is widely used to model various real-world relationships, such as e-commerce and peer review systems. Many GNN-based methods have been proposed for learning representations in such graphs. Recent approaches augment graphs by inserting new edges between nodes of the same type based on social theory to improve learning. However, they rely on a simplistic message passing design, which is prone to over-smoothing and vulnerable to noisy interactions in real-world graphs. Furthermore, they suffer from inefficiency due to their heavy design. We propose ELISE, a lightweight GNN-based method for learning node representations in signed bipartite graphs. We first extend personalized propagation to signed bipartite graphs, incorporating signed edges without adding extra edges, mitigating over-smoothing. We then jointly learn node embeddings on a low-rank approximation of the graph, reducing noise and enhancing expressiveness without sacrificing efficiency. Through extensive experiments on real-world signed bipartite graphs, we show that ELISE outperforms its competitors in link sign prediction while achieving faster training and inference.