GRATE Enhances Knowledge Graph Models for Temporal Inductive Transfer.

Current knowledge graph foundation models, like Ultra and Trix, excel at inductive transfer by learning representations that generalize to unseen entities and relations in static graphs. However, extending this capability to temporal knowledge graphs (TKGs), which involve evolving relationships over time, remains a significant challenge. Existing temporal models are often tied to specific datasets, limiting their ability to transfer to TKGs with entirely new vocabularies. Researchers propose GRATE (Gated Rotary Attention for Temporal Encoding), an innovative approach that integrates into NBFNet-style KG foundation models. GRATE uses an entity-side message function that encodes time through relative differences, rotating edge messages based on their time gap to the query and applying a query-conditioned gate to select relevant temporal signals. Crucially, it adds no new learnable parameters, thus maintaining structural transferability. To evaluate GRATE's effectiveness, new inductive transfer benchmark suites, GDELTIndT and WIKIIndT, were created. These benchmarks feature disjoint entities, relations, and timestamps, allowing for rigorous testing of cross-dataset temporal transfer, including both interpolation and extrapolation scenarios. Experiments show that a single, jointly pretrained GRATE checkpoint consistently improves performance over static base models across these challenging benchmarks and forecasting datasets.