A Structure-Feature Dynamic Decoupling GNN architecture for link prediction

Abstract

Link prediction aims to forecast the missing links within a graph, which is widely applied in various fields such as recommender systems and drug analysis. Graph Neural Networks (GNNs) have emerged as strong baselines for link prediction due to their ability to simultaneously capture the topological structure and node features of graphs. Moreover, existing approaches use the node features as the initial embedding of nodes and input them into GNNs for message passing and updating. However, these methods assume that the features and topology in graphs are homophonic and do not take into account the possible incompatibility that is common and even very severe in some graphs, harming the performance of GNNs in link prediction.

To address this issue, we propose a Structure-Feature Dynamic Decoupling GNN architecture (SFDDGNN), which mainly consists of two decoupled embedding pipelines and the dynamic gate fusion mechanism. Specifically, to avoid the incompatibility, we first utilize a GraphSAGE-based structure encoder to capture the topological structure in one pipeline. Then we construct a graph contrastive learning module to train the node feature embedding in the other pipeline. Finally, we dynamically aggregate the topology and features embedding based on the graph data distribution knowledge. Experimental results on three real-world datasets of link prediction demonstrate that SFDDGNN outperforms the state-of-the-art baselines by up to 3.54% and 6.55% in terms of AP and AUC, respectively.