Contextual Graph Attention for Knowledge Graph Reasoning

Gengchen Mai, Krzysztof Janowicz, Bo Yan, Rui Zhu, Ling Cai (STKO Lab, UCSB), Ni Lao (SayMosaic). arXiv:1910.00084v1 (September 2019).

Note: Source file named “Classification.txt” — content mismatch documented per decision D008.

Key Contribution

Proposes Contextual Graph Attention (CGA), a multi-head attention-based model for answering logical queries over incomplete knowledge graphs.

Problem

Previous KG embedding approaches for query answering either:

• Treat embedding learning and query answering as separate tasks
• Fail to handle variable contributions from different query paths

Approach

• Uses graph attention mechanism to weight different query paths
• Solves the “missing center node” problem by using initial neighborhood aggregation to generate the center embedding
• Trained jointly on original KG structure + sampled query-answer pairs
• Introduces two new datasets (DB18, WikiGeo19) with more relation types than existing benchmarks

Results

CGA with fewer learnable parameters consistently outperforms baseline models on DB18, WikiGeo19, and Bio datasets.

Connection to Vault

Extends knowledge graph reasoning from Knowledge Vault probabilistic fusion. Related to Geographic QA research (same Mai/Janowicz/Lao team). Builds on the KB inference lineage of ProPPR.