Valid for Free: Homophily-Gated Conformal Prediction for Training-Free Node Classification with Tabular Foundation Models
The study compares TabICL, a training-free tabular foundation model, against GCN+TS for node classification on graphs. TabICL showed lower expected calibration error across nine datasets. A new method, HG-DAPS, uses adjusted homophily to reduce prediction set size compared to standard methods while maintaining coverage guarantees.