Are Common Substructures Transferable? Riemannian Graph Foundation Model with Neural Vector Bundles
Researchers have introduced GAUGE, a new graph foundation model that leverages Riemannian geometry to understand transferable structures. This framework, called Neural Vector Bundle, parses intrinsic geometry using local coordinates. GAUGE is designed for pretraining and has demonstrated superior expressiveness in tasks like zero-shot link prediction and graph isomorphism. AI
IMPACT Introduces a novel geometric approach to graph foundation models, potentially improving transfer learning capabilities.