Efficient Graph Embedding Method for Link Prediction via Incorporating Graph Structure and Node Attributes

Weisheng Li, Feiyi Tang, Chao Chang, Hao Zhong, Ronghua Lin, Yong Tang

Abstract: Link prediction is a crucial task in graph analysis that aims to predict the existence of missing links in a graph. Graph embedding methods have gained popularity for link prediction by learning low-dimensional vector representations for nodes. However, most existing methods have not effectively utilized the graph structure and node attributes in attributed graphs. Additionally, higher-order neighborhood nodes that have high correlation with the source node are often disregarded. To address these issues, we propose an efficient graph embedding method called GSNA (\(\textbf{G}\)raph \(\textbf{S}\)tructure and \(\textbf{N}\)ode \(\textbf{A}\)ttributes), which effectively incorporates graph structure and node attributes to learn nodes representation for link prediction. Specifically, node sequences are generated using random walk, and then the nodes with high frequency are screened out. Then, the Top-N list of high-order neighborhood nodes with high correlation is obtained by calculating attribute similarities between nodes. Afterwards, a new graph is generated and passed through a multi-head attention mechanism to further learn the representations for link prediction tasks. Experiments conducted on several benchmark datasets demonstrate the effectiveness and efficiency of our proposed method for link prediction.