MGHA: Multi-Granularity Hierarchical Aggregation for Knowledge Graph Embedding

Knowledge Graph Embedding (KGE) involves embedding entities and relations from a knowledge graph (KG) into continuous vector spaces, learning the representations of entities and relations while preserving the inherent structure of the KGs. However, current research in KGE focuses primarily on learning global representations, with relatively little attention paid to effectively modeling complex relational heterogeneity and capturing interdependencies across different scales. This results in challenges when attempting to model multi-level, cross-scale feature fusion. To address this issue, we propose Multi-Granularity Hierarchical Aggregation for Knowledge Graph Embedding (MGHA). This method employs a multi-granularity hierarchical encoder for global context extraction and local structure encoding, enabling feature modeling from the global to the local level and effectively capturing cross-scale information. Building upon this, the model introduces a channel attention mechanism to adaptively weight the aggregated features, thereby achieving hierarchical fusion of multi-granularity information and ultimately generating embedding features that simultaneously represent both global and local information. Experimental results demonstrate that MGHA achieves competitive results on four widely-used benchmark datasets.