Reproduce “Scalable Graph Neural Networks via Bidirectional Propagation” 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
Ming Chen, Zhewei Wei, Bolin Ding, Yaliang Li, Ye Yuan, Xiaoyong Du, Ji-Rong Wen
GBP (Graph neural network via Bidirectional Propagation) introduces a scalable GNN that achieves sub-linear time complexity for precomputation, training, and inference. The method successfully processes graphs with billions of edges on a single machine, demonstrating superior accuracy and up to 10x faster performance compared to leading baselines on large datasets.
Key insights
- A combined Monte-Carlo sampling and deterministic push propagation strategy can create an unbiased estimator for the feature propagation matrix while effectively reducing variance for large graphs.
- Achieving sub-linear time complexity for GNNs requires decoupling feature propagation from neural network training and optimizing the propagation step itself.
- Careful parameter tuning for the number of random walks and push threshold allows balancing computational costs, leading to an optimal theoretical time complexity for GNN precomputation, training, and inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.