Attention and Time Perception Based Link Prediction in Dynamic Networks

There are numerous applications that could be modeled as networks and predicting their relationships within their evolution process is an important task. The continuous changing of the network makes link prediction challenge, but its structure and temporal information will give clues to support this task. Inspired by this, we proposed a novel link prediction model that incorporates attention and time-awareness mechanisms to extract the latent pattern from the spatial and temporal information. A Transformer model is introduced to capture the long-time dependencies between nodes, combined with link similarity coding to capture the correlation between links in dynamic networks, while neighbour weights are calculated based on temporal differences to aggregate neighbour information. Ultimately, a classifier is employed for link prediction. Comprehensive experiments on real-world datasets demonstrate the superior performance of proposed model than other baselines.