Researchers have introduced GRATE (Gated Rotary Attention for Temporal Encoding), a novel method designed to enhance the temporal transferability of knowledge graph foundation models. GRATE operates by adding no new learnable parameters and encoding time through relative differences, using a query-conditioned gate to select relevant temporal signals. This approach integrates seamlessly with existing NBFNet-style models while maintaining structural transferability. To evaluate GRATE's performance on inductive transfer tasks with disjoint vocabularies, new benchmark suites named GDELTIndT and WIKIIndT were developed. AI
IMPACT Enhances temporal reasoning capabilities in knowledge graph foundation models, potentially improving applications that rely on time-sensitive data.
RANK_REASON The cluster contains an academic paper detailing a new method for knowledge graph foundation models. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Gated Rotary Attention for Temporal Encoding
- GDELTIndT
- GRATE
- Knowledge graph foundation models
- NBFNet
- Temporal knowledge graphs
- Trix
- WIKIIndT
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
