Streaming Dynamic Graph Neural Networks for Continuous-Time Temporal Graph Modeling

Abstract

Dynamic graphs are suitable for modeling structured data that evolve over time and have been widely used in many application scenarios such as social networks, financial transaction networks, and recommendation systems. Recently, many dynamic graph methods are proposed to deal with temporal networks. However, due to the limitations of storage space and computational efficiency, most approaches evolve node representations by aggregating the latest state information of neighbor nodes, thus losing a lot of information about neighbor nodes’ state changes. Besides, high computational complexity makes it challenging to deploy dynamic graph algorithms in real-time. To tackle these challenges, we propose a novel streaming dynamic graph neural network (SDGNN) for modeling continuous-time temporal graphs, which can fully capture the state changes of neighbors and reduce the computational complexity of inference. Under SDGNN, an incremental update component is designed to incrementally update node representation based on the interaction sequence, an inference component is utilized for specific downstream tasks, and a message propagation component is employed to propagate interactive information to the influenced nodes by considering the update time interval, position distance, and influence strengths. Extensive experiments demonstrated that the proposed approach significantly outperforms state-of-the-art methods by capturing more state change information and efficient parallelization.