Reifying input graphs into a fixed meta-relation vocabulary allows vanilla GNNs to achieve zero-shot link prediction on unseen graphs. A standard GAT trained on a single small graph matches the performance of the specialized ULTRA foundation model across 40 inductive benchmarks.
HOW THIS AFFECTS YOU
●
researcherYou can achieve foundation-model performance on knowledge graphs using simple architectures and standardized representations.