{"id":"arxiv/2601.23239","url":"https://arxiv.org/abs/2601.23239","tags":["stat.ML","cs.IT","cs.LG","cs.SI","math.ST"],"score":0,"title":"Graph Attention Network for Node Regression on Random Geometric Graphs with Erdős--Rényi contamination","domain":"stat.ML","source":"arxiv","authors":["Somak Laha","Suqi Liu","Morgane Austern"],"reviews":0,"subject":"stat.ML","summary":"Graph attention networks (GATs) are widely used and often appear robust to noise in node covariates and edges, yet rigorous statistical guarantees demonstrating a provable advantage of GATs over non-attention graph neural networks~(GNNs) are scarce. We partially address this gap for node regression with graph-based errors-in-variables models under simultaneous covariate and edge corruption: responses are generated from latent node-level covariates, but only noise-perturbed versions of the latent covariates are observed; and the sample graph is a random geometric graph created from the node…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-01-30","reactions":{}}