Dynamic Graph Neural Networks (Dynamic GNNs) have emerged as powerful tools for modeling real-world networks with evolving topologies and node attributes over time. A survey by Professors Zhewei Wei, Yanping Zheng, and Lu Yi at Renmin University of China, published on 15 June 2025 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature, offers an extensive review of Dynamic GNNs. It covers the foundational knowledge, prevalent models in the current landscape, and outlines prospective research avenues and technological advancements.
Based on the granularity of time step, this review categorizes dynamic graphs into Continuous-Time Dynamic Graphs (CTDGs) and Discrete-Time Dynamic Graphs (DTDGs). Each event in a CTDG has a timestamp, which smoothly tracks the evolution of the graph structure and allows for more accurate analysis of temporal relations and dependencies between events. DTDGs, in contrast, decompose the dynamic graph into a series of static networks, capturing the graph structure at various selected time points.

Adhering to the taxonomy above, the researchers conduct a comprehensive review of dynamic GNN models, offering a thorough assessment of their strengths and weaknesses. They streamline the diverse models of CTDGs and DTDGs into unified architectural categories, with DTDGs being subdivided into stacked or integrated architectures. The survey also covers ongoing research trends, such as models and systems designed for large-scale dynamic graphs, and explores potential future research directions.
DOI
10.1007/s11704-024-3853-2