Reproduce “Non-exchangeable Conformal Prediction for Temporal 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
Tuo Wang, Jian Kang, Yujun Yan, Adithya Kulkarni, Dawei Zhou
NCPNet, a framework developed at Virginia Tech and collaborating institutions, introduces a novel approach for quantifying prediction uncertainty in Temporal Graph Neural Networks (TGNs) by explicitly addressing non-exchangeability. It achieves guaranteed coverage while reducing prediction set size by up to 34.2% on real-world temporal graphs compared to baselines.
Key insights
- Temporal dependencies in dynamic graphs fundamentally violate the exchangeability assumption of standard Conformal Prediction, requiring explicit theoretical and methodological compensation.
- Propagating uncertainty through both topological and temporal diffusion within a non-conformity score effectively captures the complex, evolving dependencies in temporal graphs.
- Optimizing weighted parameters in quantile calculation is crucial for simultaneously achieving desired statistical coverage and minimizing the size of prediction sets in non-exchangeable temporal contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.