Streaming Graph Neural Networks via Generative Replay

Recent advances in graph neural networks (GNNs) have achieved strong performance on static graph tasks, but real‑world graphs often evolve as streams, making continual updates essential. Retraining on the whole graph each time is computationally prohibitive, while naive online learning suffers from catastrophic forgetting.

We propose a novel framework that couples a primary GNN with an auxiliary generative graph model. The generator, built on a GAN architecture, learns to reproduce the distribution of historical graph structures without storing raw data, while the discriminator distinguishes real from synthetic random‑walk sequences. The GNN consumes both newly observed nodes and replayed synthetic nodes, using soft labels from the previous GNN to preserve past knowledge.

To adapt to evolving graphs, we define incremental updates for both models: new patterns are incorporated by training on changed neighborhoods, and obsolete patterns are filtered out via a replay‑node pruning strategy. This design enables the system to learn new information and retain old knowledge simultaneously.

Experiments on four public streaming graph datasets show that our method (SGNN‑GR) surpasses existing incremental GNN approaches and approaches the performance of fully retrained models, while being 3–6× faster per iteration and 1.5–2× faster overall.

We conclude that generative replay offers an effective solution for continual learning in streaming graph scenarios and outline future research directions such as adaptive model expansion, unsupervised continual GNNs, and interpretability.