Multi-relation constrained co-occurrence attention model for knowledge graph representation learning

doi:10.1016/j.neucom.2026.135181

Multi-relation constrained co-occurrence attention model for knowledge graph representation learning

2026-09-19

0

PRE

AI

DOI：10.1016/j.neucom.2026.135181

Original

Original request for help

Save

Abstract

En 中文

• A novel hybrid framework (MRCCA) synergizes Graph Attention Networks (GATs) and Knowledge Graph Embedding (KGE) to overcome over-smoothing and limited relational capacity in Knowledge Graph Completion. • A co-occurrence-driven attention mechanism exploits co-occurrence information and KGE semantic assumptions, eliminating conventional linear transformations and nonlinear activations to reduce trainable parameters. • A Multi-Relational Matching Module incorporating KGE geometric operations (rotation and translation) to explicitly model complex relations such as one-to-many and many-to-one mappings, across multiple evaluation metrics, achieving high parameter efficiency, yielding state-of-the-art results on the general benchmarks (FB15k-237, WN18RR) and the domain specific KGs (Kinship, UMLS).

Keywords:

Link prediction

Knowledge graph embedding

Graph neural networks

Co-occurrence attention

Attention mechanism

Multi-relation constrained co-occurrence attention model for knowledge graph representation learning

Abstract

Journal

Organization

Cited Papers