SEAL+: A subgraph-enhanced framework for link prediction with graph neural networks

Abstract

Link prediction is a critical research topic in network analysis, typically formulated as a classification problem where the goal is to determine whether a link exists between a pair of nodes (denoted as 1 in existence and 0 for non-existence). In most existing works, the feature vectors of a pair of nodes are combined to obtain the feature vector representing the link between them; these feature vectors (or embeddings) are constructed using graph neural networks (GNNs). This paper uses a GNN-based link prediction method called EAL as the baseline. EAL consists of an Encoder and a Decoder. A significant challenge in EAL is that the feature vector extracted by the GNN can be identical for different pairs of nodes. To address this issue, we propose leveraging the concept of subgraphs to enhance link prediction performance. To this end, the Encoder is equipped with subgraphs, forming the SEAL framework. One limitation of SEAL is that it generates identical link representations for different links when the embeddings of the nodes involved are the same. To overcome this limitation, the Decoder of SEAL also uses the subgraph information, resulting in the novel framework SEAL+. We evaluate these two frameworks against baseline methods using various metrics, demonstrating their superiority. Specifically, SEAL+ achieves average improvements of 10.25 %, 17.25 %, 3.75 %, and 4.65 % in terms of accuracy, F1-Score, average precision, and area under the precision-recall curve, respectively, compared to the SEAL.