Reproduce “Graph Structure Learning with Variational Information Bottleneck” 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
Qingyun Sun, Jianxin Li, Hao Peng, Jia Wu, Xingcheng Fu, Cheng Ji, Philip S. Yu
VIB-GSL learns task-specific graphs and compressed representations for graph classification. It masks node-feature components and infers a sparse graph from features rather than relying on observed links, balancing label prediction with compression. On REDDIT-B with a GIN backbone, it reaches 88.5% ± 1.8 accuracy, versus NeuralSparse’s 86.2% ± 2.7. Under synthetic edge attacks, accuracy stays comparatively level, tested on one benchmark.
Key insights
- The Information Bottleneck principle provides a theoretically sound and universal approach to distill essential, task-relevant information for graph structure learning, making the resulting graph robust to nuisances.
- Optimal graph structures for GNNs are not always the raw observed graphs but rather compressed, task-specific representations that balance predictive power with data compression.
- The 'beta' parameter in the IB objective controls the trade-off between prediction and compression, and its optimal value is dataset-dependent, influencing the density and informativeness of the learned graph.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.