1138 - GraphOracle: Efficient Fully-Inductive Knowledge Graph Reasoning via Relation-Dependency Graphs
Association for Artificial Intelligence 2026, Yongqi Zhang, Siyi Liu, Enjun Du
Open MIND
Problems Identified (4)
Fully-inductive KG reasoning: Knowledge graph reasoning is challenging when both test-time entities and relations are unseen during training.
Cross-domain KG generalization: The work addresses reasoning performance in cross-domain scenarios.
Fully-inductive KG reasoning: Knowledge graph reasoning is challenging when both test-time entities and relations are unseen during training.
Cross-domain KG generalization: The work addresses reasoning performance in cross-domain scenarios.
Proposed Solutions (5)
GraphOracle RDG framework: GraphOracle transforms each knowledge graph into a Relation-Dependency Graph to support robust fully-inductive reasoning.
Relation-dependency graph encoding: The Relation-Dependency Graph encodes directed precedence links between relations to capture compositional patterns while reducing graph density.
Query-conditioned attention propagation: A multi-head attention mechanism conditioned on the query relation propagates information over the RDG to create context-aware relation embeddings.
Relation-guided inductive GNN: The generated relation embeddings guide a second GNN to perform inductive message passing over the original knowledge graph.
GraphOracle RDG framework: GraphOracle transforms each knowledge graph into a Relation-Dependency Graph to support robust fully-inductive reasoning.
Results (3)
Improved fully-inductive performance:
Improved cross-domain performance:
RDG attention ablation support:
Research Domain
knowledge graph reasoning