SAMI: A Structure-Aware Multi-Partition Embedding Interaction Model for Accurate Link Prediction in Knowledge Graphs

Knowledge graph embedding (KGE), which applies representation learning to represent entities and relationships in knowledge graphs, has attracted significant attention from researchers due to its potential applications in various domains. However, most of the existing KGE methods suffer from the limitation of using single semantic information. This limitation fails to capture the complex and structural information in knowledge graphs (KGs). In this paper, we introduce a novel method called the Structure-Aware Enhanced Multi- Partition Embedding Interaction (SAMI) model for Knowledge Graph Embedding (KGE). SAMI leverages both graph attention network and tensor decomposition to learn expressive and structural enhanced representations for KGs. Specifically, it uses the graph attention layers to aggregate nodes' features in a neighborhood as an encoder and utilizes an Enhanced Multi- Partition Embedding Interaction (EMEI) to learn independent local features as a decoder. SAMI shows impressive results on several popular datasets compared with baseline methods in terms of both Mean Reciprocal Rank (MRR) and Hits@K.