Learning Temporal Graph Representation via Memory-Aware Autoencoder

Representation learning over temporal graphs has garnered significant attention from researchers due to the prevalence of real-world graphs that naturally evolve over time. However, some existing methods that discretize temporal graphs into snapshots often result in the loss of temporal information. On the other hand, some continuous-time methods ensure temporal continuity but neglect the intrinsic correlations among events and the influence of node features on their behavior. Therefore, we propose a novel temporal graph representation learning model with memory-aware autoencoder, called MATE, which consists of a node-wise memory unit and a temporal representation learning module. The temporal representation learning module is on the basis of autoencoder. Specifically, a temporal GNN-based encoder is constructed to aggregate temporal topological neighborhood features, while a dynamic event decoder is employed to learn the dynamic evolution of the temporal graph. Additionally, a static feature decoder is included to guide the encoder to maximize the retention of node features in its temporal representation. As for the node-wise memory unit, MATE utilizes it to store the historical information of nodes. Through the interactive iteration of the node-wise memory unit and temporal representation learning module, MATE is able to capture the intrinsic correlations among events occurring in a sequence. Extensive experiments conducted on various real-world datasets demonstrate the superiority of MATE.