Summary: | Abstract Dynamic network link prediction has practical applications in many areas, such as social networks, traffic networks, biological networks, and citation networks. Because of its essential practical significance, it has attracted the attention of many researchers. The key to dynamic network link prediction is to model the network topology evolution and capture time information. Currently, most studies divide dynamic networks into a series of static time snapshots, which can be considered a rough compressed of continuous-time dynamic networks. Such compression will lead to the loss of time evolution information in the window and how to choose the appropriate partition granularity is a considerable challenge. In this paper, we propose a link prediction method for continuous-time networks based on Random Walk and Time Aggregation(RWTA). In the method, we perform random walk with time-constrained directly on a continuous-time network to get node sequence without slicing into time snapshots. Then, based on skip-gram, the initial node representation is gotten, and the dynamic graph with node representation is created. A temporal proximity neighborhood aggregation process is designed to enhance node representation, and the binary operator is done to obtain edge representation. Finally, A classifier is utilized to predict links. Extensive experiments on real-world datasets show that our model outperforms other state-of-the-art methods.
|