Generative Graph Pattern Machine

Tianyi Ma (University of Notre Dame) · Zheyuan Zhang (University of Notre Dame) · Zehong Wang (University of Notre Dame) · Chuxu Zhang (University of Connecticut) · Yanfang Ye (University of Notre Dame)

architectural choicescross-graph pretrainingexpressiveness generalizable representationsgenerative graph pattern machinegenerative transformergraph neural networkslocal neighborhood aggregationlong-range dependencies message-passing node representations over-smoothing over-squashingpre-training frameworkscalability transfer learning

Graph neural networks (GNNs) have been predominantly driven by message-passing, where node representations are iteratively updated via local neighborhood aggregation. Despite their success, message-passing suffers from fundamental limitations