Research map / Graph learning / Knowledge and temporal graphs

Knowledge Graph Completion with Relation-Aware Anchor Enhancement

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Inductive knowledge graph completion · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RAA-KGC, a text-based knowledge graph completion method that samples relation-aware neighbor entities of the head entity as anchors and pulls query embeddings toward them, built on SimKGC.

Problem

Text-based knowledge graph completion via bi-encoder contrastive learning lacks context about the target entity, so queries (h, r, ?) carry little information about what the answer looks like.

Approach

Use the tail entities connected to the query head entity by the same relation (relation-aware entities) as anchors, append their descriptions to the query to form anchor-enhanced queries, and train with an extra InfoNCE loss so the query embedding is pulled toward the anchor neighborhood; inference sums the anchor-enhanced and classic scores.

Key results

• On WN18RR, RAA-KGC reaches MRR 59.74 and Hit@10 76.01 versus SimKGC's 55.31 and 73.65.
• On Wikidata5M-Trans, RAA-KGC reaches MRR 34.15 versus SimKGC's 29.18; on FB15k-237 it is only 31.46 vs 29.89 MRR but it trails RotatE's Hit@1/Hit@3 in the table.
• Plugging the anchor mechanism into SimKGC raises WN18RR_v1 MRR from 56.23 to 63.21 (per the compatibility table).

Measured effects reported in the paper

Surprising findings

• The method falls short on FB15k-237 relative to some triple-based baselines, which the authors attribute to that dataset's denser topology.
• Sample size 5 of anchors performs slightly worse than 4, which the authors attribute to noise from too many relation-aware entities.

Limitations

• Smaller gains on FB15k-237, a denser graph, where the method underperforms some triple-based baselines.
• Performance degrades with too many anchors (sample size 5), attributed to noise.
• Relies on relation-aware neighbors existing; results for relations like member_of_domain_region are worse than SimKGC.

Experimental setup

Datasets: WN18RR FB15k-237 Wikidata5M-Trans

Benchmarks: WN18RR FB15k-237 Wikidata5M-Trans WN18RR_v1-v4 inductive FB15k-237_v1-v4 inductive

Metrics: MRR Hit@1 Hit@3 Hit@10

Baselines: TransE ComplEx DistMult ConvE RGCN RotatE QuatE KG-BERT MTL-KGC StAR SimKGC HaSa

Builds on: SimKGC KG-BERT InsKGC

Closest papers by meaning

• Select and Augment: Enhanced Dense Retrieval Knowledge Graph Augmentation (Abstract Reprint) (AAAI 2024 · similarity 0.91)
• Anchoring Path for Inductive Relation Prediction in Knowledge Graphs (AAAI 2024 · similarity 0.89)
• Clustering then Propagation: Select Better Anchors for Knowledge Graph Embedding (NeurIPS 2024 · similarity 0.89)
• Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.89)
• Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 2025 · similarity 0.89)
• Revisiting Document-Level Relation Extraction with Context-Guided Link Prediction (AAAI 2024 · similarity 0.88)