Homophilous-graph stratum coverage under HG-DAPS

Determine whether Homophily-Gated Diffusion Adaptive Prediction Sets (HG-DAPS) lowers coverage within homophilous graph-node strata when its adjusted-homophily gate activates score diffusion.

Background

The paper introduces HG-DAPS, a training-free conformal prediction method that diffuses nonconformity scores over graph edges using a diffusion weight determined by adjusted homophily computed from labeled context nodes. Its finite-sample guarantee establishes marginal coverage, but the experiments show that marginal coverage can conceal undercoverage in low-homophily subgroups, particularly when raw rather than adjusted homophily is used for gating on class-imbalanced heterophilous graphs.

The reported stratum-coverage analysis is limited to two heterophilous trap graphs. Although HG-DAPS reduces prediction-set size on homophilous graphs and activates diffusion there, the paper does not determine whether this diffusion causes lower coverage for particular node strata within homophilous graphs.