Reproduce “Streaming Graph Neural Networks via Continual Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Junshan Wang, Guojie Song, Yi Wu, Liang Wang

Researchers from Peking University and Alibaba Group developed ContinualGNN, a framework that enables Graph Neural Networks to adapt efficiently to evolving network data. The method integrates continual learning principles to detect new patterns and preserve historical knowledge, achieving classification performance comparable to full retraining while being significantly more computationally efficient and mitigating catastrophic forgetting.

Key insights

  • Combining data replaying with parameter regularization based on Fisher information effectively counters catastrophic forgetting in graph contexts.
  • Approximating influence propagation in GNNs via linearized aggregation and Breadth-First Search (BFS) significantly speeds up the detection of nodes affected by network changes.
  • A hierarchy-importance sampling strategy improves the representativeness of historical data for memory-based replay in dynamic graph settings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.