Diffusion Model for Continuous-Time Dynamic Graph Generation

Dynamic graph generation has emerged as a powerful paradigm for modeling evolving structural phenomena. By synthesizing a stream of events, it can capture real-world network evolution over time. However, generation remains limited and slow as dynamic graphs grow in both size and feature dimensionality, limiting practical use. We introduce Dynamic Graph Diffusion (DyGDi), a new dynamic graph generation model that leverages diffusion to jointly model a window of events and their features. DyGDi generates the entire window in a single pass. This yields faster generation and lets events and features attend to one another, allowing the model to learn more coherent dynamics. To address a node-identity issue arising from this parallel generation, we introduce a new event tokenization that injects more information about event appearance, anchoring newly generated nodes to their identities. Altogether, we show that DyGDi trains and generates more efficiently than previous models and scales better. It produces graphs with high-quality node-feature and topological properties, while adapting to zero-shot settings such as link prediction.