GRAPHLCP: Structure-Aware Conformal Prediction Gives Distribution-Free Uncertainty on Graphs

arXiv·low signalXBlueskyLinkedInCopy link

Extends conformal prediction to graph-structured data by incorporating local graph structure, providing finite-sample coverage guarantees without distributional assumptions. Standard conformal prediction ignores graph topology and produces overly wide prediction sets; GRAPHLCP localizes predictions using neighborhood structure. Useful for GNN practitioners who need reliable uncertainty quantification in production (fraud detection, molecular property prediction).