Information bottleneck for link prediction

Link prediction is a fundamental problem in graph structures with a wide range of applications in the real world. Graph neural networks (GNNs) have shown impressive performance in predicting links by capturing information within the graph. However, a single GNN model often struggles to capture all the relevant information and may exhibit noticeable differences when applied to datasets of varying sizes. In addition, how different types of information communicate and collaborate to jointly influence downstream tasks is also an important challenge. To address this, we propose a novel model LGnet (Local and Global Net) that can simultaneously capture both local and global information, efficiently and seamlessly integrating them. This model employs two distinct approaches to capture local and global information. Local information is captured through operations on neighboring nodes, where the node interactions are modeled using a pairwise method combined with attention mechanisms. Meanwhile, global information is obtained from the power subgraphs of the graph. These two types of information are then processed through an information bottleneck, which selects the most critical features for making predictions. Experiments show that LGnet outperforms existing methods in several real-world datasets.