Research map / Graph learning / Knowledge and temporal graphs

KGCRR: An Effective Metric-Driven Knowledge Graph Completion Framework by Designing a Novel Upper Bound Function with Adaptive Approximation to Reciprocal Rank

AAAI 2025 · Graph learning › Knowledge and temporal graphs · approach: Inductive knowledge graph completion · 1 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 CRR, a log-transformed sigmoid-based upper bound on reciprocal rank with a pressure parameter ρ, used as a plug-and-play loss for knowledge graph embedding models on FB15k-237 and WN18RR.

Problem

Knowledge graph completion is typically trained with margin-based or log-likelihood losses that are misaligned with the reciprocal rank (RR/MRR) evaluation metric, while directly optimizing RR suffers from gradient vanishing.

Approach

CRR is a nonlinear (log) transformation of a rank-based surrogate built on a shifted sigmoid with pressure parameter ρ; increasing ρ makes the surrogate a closer approximation to RR, and the loss is a plug-and-play replacement for existing KGE objectives.

Key results

• CRR loss improves MRR over margin-based loss by an average of 26.49% on FB15k-237 and 7.51% on WN18RR, and over log-likelihood loss by 32.31% and 9.93% respectively.
• Removing the log transformation (w/o T) lowers CRR performance across models, e.g., TransE on WN18RR drops from 0.241 to 0.207 MRR.
• Across the two datasets and seven KGE models, the paper reports an average 19.06% MRR improvement from CRR.

Measured effects reported in the paper

Surprising findings

• The best loss depends on dataset and model: ComplEx with log-likelihood loss outperforms margin-based loss on WN18RR but not on FB15k-237.
• Increasing the pressure parameter ρ beyond an optimum degrades performance, likely due to gradient vanishing similar to direct RR loss.

Limitations

• Gains are larger on FB15k-237 than on WN18RR, possibly because of fewer candidate entities per query.
• Only the log function was tested as the transformation; other transformations were left for future work.
• Evaluated only on two standard datasets with shallow KGE models and pressure/temperature hyperparameters tuned by grid search.

Experimental setup

Datasets: FB15k-237 WN18RR

Benchmarks: FB15k-237 WN18RR

Metrics: MRR Hits@1 Hits@10

Baselines: margin-based loss log-likelihood loss rank loss RR loss PairRE CompoundE

Builds on: Smooth-AP Smooth-AUC OpenKE

Closest papers by meaning

• Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 2025 · similarity 0.88)
• Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.88)
• Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.87)
• KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge (NeurIPS 2024 · similarity 0.87)
• Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.87)
• SPAC: Sparse Partitioning and Adaptive Core Tensor Pruning Model for Knowledge Graph Completion (AAAI 2025 · similarity 0.87)