Reproduce “Uncertainty Quantification over Graph with Conformalized Graph Neural 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

Kexin Huang, Ying Jin, Emmanuel Candès, Jure Leskovec

The Conformalized GNN (CF-GNN) framework provides statistically guaranteed uncertainty quantification for Graph Neural Networks in transductive settings. It achieves this by ensuring marginal coverage and significantly reducing prediction set sizes by up to 74% for classification and around 7% for regression, compared to vanilla conformal prediction.

Key insights

  • Conformal Prediction can provide valid marginal coverage for GNNs on graph data under the condition of permutation invariance of the non-conformity score function, effectively extending CP to non-i.i.d. graph structures.
  • The inefficiency of conformal predictions on graphs is often correlated along network edges, indicating that local graph topology can be leveraged to optimize prediction set/interval sizes.
  • A learnable, topology-aware correction mechanism, trained with a conformal-aware inefficiency loss, can systematically improve the efficiency of uncertainty quantification without sacrificing statistical guarantees.

Summary by alphaXiv.

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