Research map / Graph learning / Knowledge and temporal graphs

Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Inductive knowledge graph completion · 4 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 CBLiP, a Transformer-only subgraph encoder for inductive and transductive KG completion that uses connection-biased attention and entity role embeddings in place of explicit path encoding.

Problem

Inductive knowledge graph completion, where test-time entities are unseen during training; top-performing models rely on costly explicit path encoding modules with extra hyperparameters.

Approach

Adapts GRAN's connection-biased attention to subgraph triples: a connection-type matrix (shared heads/tails, inverse, parallel edges, none) adds learned key and value biases in Transformer attention, and entity role embeddings (head, tail, other) replace relative-distance labeling, so paths are implicitly captured without path extraction.

Key results

• CBLiP achieves best Hits@10 in 7 of 12 inductive splits, e.g., 97.30 on WN18RR v1 versus 94.80 for NBFNet.
• On inductive FB15K-237 splits, CBLiP trails path-based NBFNet (e.g., 80.10 vs 96.00 on v4), which the authors attribute to dense neighborhoods.
• On transductive relation prediction, CBLiP reaches MRR 0.976 on WN18RR, above PathCon's 0.974, and MRR 0.919 on NELL995 versus PathCon's 0.896.

Measured effects reported in the paper

Surprising findings

• CBLiP underperforms on FB15K-237 in the transductive relation prediction setting (MRR 0.863 vs PathCon 0.984), which the authors attribute to high triple-to-entity ratio.

Limitations

• Underperforms path-based models on FB15K-237 inductive and transductive splits with dense neighborhoods.
• Only addresses inductive entity prediction and transductive relation prediction; fully inductive setting with unseen relations is future work.

Experimental setup

Datasets: WN18RR inductive splits v1-v4 FB15K-237 inductive splits v1-v4 NELL-995 inductive splits v1-v4 WN18RR FB15K-237 NELL995

Benchmarks: GraIL inductive splits (WN18RR, FB15K-237, NELL-995) WN18RR, FB15K-237, NELL995 transductive relation prediction WN18, DDB14, FB15K (appendix)

Metrics: Hits@10 MRR Hits@1 Hits@3

Baselines: NeuralLP DRUM RuleN GraIL CoMPILE TACT ConGLR Report LogCo NBFNet PathCon

Builds on: GRAN GraIL Transformer

Closest papers by meaning

• Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.91)
• Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.89)
• Anchoring Path for Inductive Relation Prediction in Knowledge Graphs (AAAI 2024 · similarity 0.89)
• S²DN: Learning to Denoise Unconvincing Knowledge for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.89)
• Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.89)
• KGCRR: An Effective Metric-Driven Knowledge Graph Completion Framework by Designing a Novel Upper Bound Function with Adaptive Approximation to Reciprocal Rank (AAAI 2025 · similarity 0.88)