Multi-granularity Evolution Network for Dynamic Link Prediction

Dynamic link prediction target to predict future new links in a dynamic network, is widely used in social networks, knowledge graphs, etc. Some existing dynamic methods capture structural characteristics and learn the evolution process from the entire graph, which pays no attention to the association between subgraphs and ignores that graphs under different granularity have different evolve patterns. Although some static methods use multi-granularity subgraphs, they can hardly be applied to dynamic graphs. We propose a novel Temporal K-truss based Recurrent Graph Convolutional Network (TKRGCN) for dynamic link prediction, which learns graph embedding from different granularity subgraphs. Specifically, we employ k-truss decomposition to extract multi-granularity subgraphs which preserve both local and global structure information. Then we design a RNN framework to learn spatio-temporal graph embedding under different granularities. Extensive experiments demonstrate the effectiveness of our proposed TKRGCN and its superiority over some state-of-the-art dynamic link prediction algorithms.