Temporality- and Frequency-aware Graph Contrastive Learning for Temporal Network

Shiyin Tan, Jingyi You, Dongyuan Li

Abstract: Graph contrastive learning (GCL) methods aim to learn more distinguishable representations by contrasting positive and negative samples. They have received increasing attention in recent years due to their wide application in recommender systems and knowledge graphs. However, almost all GCL methods are applied to static networks and can not be extended to temporal networks directly. Furthermore, recent GCL models treat low- and high-frequency nodes equally in overall training objectives, which hinders the prediction precision. To solve the aforementioned problems, in this paper, we propose a <u>T</u>emporality- and <u>F</u>requency-aware <u>G</u>raph <u>C</u>ontrastive <u>L</u>earning for temporal networks (TF-GCL). Specifically, to learn more diverse representations for infrequent nodes and fully explore temporal information, we first generate two augmented views from the input graph based on topological and temporal perspectives. We then design a temporality and frequency-aware objective function to maximize the agreement between node representations of the two views. Experimental results demonstrate that TF-GCL remarkably achieves more robust node representations and significantly outperforms the state-of-the-art methods on six temporal link prediction benchmark datasets. Considering the reproducibility, we release our code on Github.