A Paper Citation Link Prediction Method Using Graph Attention Network

Link prediction utilizes existing data to prediction the relational ties between any pair of nodes within a relationship network, and then holds the potential to recover or anticipate absent relationship data. Nevertheless, the existing methods encounter the problems of performance degradation, insufficient utilization of attribute information of nodes or edges caused by the increase of network size, which brings extremely adverse effects to the citation prediction of papers. We proposed a novel approach to paper citation prediction using graph attention network (GAT), which uses a dual attention mechanism, amalgamating both the structural and content-related facets of network nodes, to adeptly gauge the influence exerted by neighboring nodes upon the central node. Additionally, it adeptly integrates attribute data associated with network nodes and edges, optimizing the graph attention network for the specific task of link prediction. The experimental results shown that our method can achieve a significant improvement in the prediction accuracy of the paper citation link on the Cora dataset.