Reproduce “Graph Belief Propagation Networks” 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

Junteng Jia, Cenk Baykal, Vamsi K. Potluru, Austin R. Benson

Graph Belief Propagation Networks (GBPNs) integrate the statistical rigor of Collective Classification with the scalability of Graph Neural Networks for semi-supervised node classification, modeling node labels as a Conditional Random Field with learned self-potentials and interpretable coupling potentials. The model demonstrates superior accuracy on synthetic MRF data and competitive to superior performance on empirical datasets, notably achieving a 3.8% F1 score improvement over baselines on the 'Payments' financial dataset and effectively handling heterophilous graphs.

Key insights

  • Explicitly modeling label correlations via a CRF and learning a coupling matrix enables the model to effectively capture both homophilous and heterophilous relational patterns in graphs.
  • Backpropagation through a fixed number of belief propagation iterations allows for end-to-end learning of both feature-based node potentials and inter-class edge potentials, bridging deep learning with probabilistic graphical models.
  • The learned coupling matrix provides direct interpretability, offering insights into the affinities between different classes within a graph, which is often missing in standard GNNs.

Summary by alphaXiv.

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