Time-Aware Graph Learning for Link Prediction on Temporal Networks.

Link prediction on temporal networks aims to predict the future edges by modeling the dynamic evolution involved in the graph data. Previous methods relying on the node/edge attributes or the distance on the graph structure are not practical due to the deficiency of the attributes and the limitation of the explicit distance estimation, respectively. Moreover, the existing graph representation learning methods mostly rely on graph neural networks (GNNs), which cannot adequately take the dynamic correlations between nodes into consideration, leading to the generating of inferior node embeddings. Thus, we propose a time-aware graph (TAG) learning method for link prediction on temporal networks. We first conduct a theoretical causal analysis proving that the correlations between nodes are required to be unchanged for the temporal graph representation learning using GNNs. Then, we model the recent dynamic node correlations by designing an edge-dropping (ED) module and adopting a recent neighbor sampling (RNS) strategy so as to approximate the above condition. Besides, we also preserve the long-term stable node correlations by introducing additional self-supervisions using the contrastive learning. Comprehensive experiments were conducted on four public temporal network datasets, i.e., MathOverflow, StackOverflow, AskUbuntu, and SuperUser, demonstrate that TAG can achieve state-of-the-art performance in terms of average precision (AP) and area under the ROC curve (AUC). In addition, TAG can ensure high computational efficiency by making the temporal graph lightweight, letting it be practical in real-world applications.