Decoupled Graph Neural Networks for Large Dynamic Graphs

Real-world graphs, such as social networks, financial transactions, and recommendation systems, often demonstrate dynamic behavior. This phenomenon, known graph stream, involves the changes of nodes emergence disappearance edges. To effectively capture both structural temporal aspects these neural networks have been developed. However, existing methods are usually tailored to process either continuous-time or discrete-time cannot be generalized from one other. In this paper, we propose a decoupled network for large including unified propagation that supports efficient computation continuous discrete graphs. Since structure-related computations only performed during process, prediction downstream task can trained separately without expensive computations, therefore any sequence model plugged-in used. As result, our algorithm achieves exceptional scalability expressiveness. We evaluate on seven real-world datasets The experimental results state-of-the-art performance in kinds Most notably, is well illustrated by its successful application graphs with up over billion edges hundred million nodes.