Dynamic graph contrastive learning based on learnable view generators

Dynamic graph representation learning aims to capture graphs' evolving structure and properties by learning embedded representations of nodes and edges over time. Graph self-supervised learning, particularly graph contrastive learning, has garnered significant attention recently. However, existing dynamic graph contrastive learning methods are vulnerable to data distribution imbalances, leading to model overfitting on active nodes and edges, and they are also prone to capturing meaningless dynamic changes caused by random noise, which degrades the performance and robustness of the model. To alleviate these problems, this paper introduces a Learnable Dynamic Graph Contrastive (LDGC) framework based on a learnable view generator. LDGC combines the learnable graph view generator with an adaptive contrastive learning mechanism, effectively reducing time-related noise in dynamic networks and mitigating data distribution imbalances. First, LDGC constructs two augmented views by embedding the learnable view generator into automatic augmentation strategies, where each graph view generator is trained to produce a distribution of graphs based on the input conditions. Then, a more discriminative dynamic node representation is learned by incorporating both time and node activity information, maximizing the consistency between the node representations of two views during the contrastive learning process. Comprehensive experiments on benchmark datasets for dynamic link prediction validate the model's effectiveness and demonstrate its superiority over current state-of-the-art methods.