Entity-relation aggregation mechanism graph neural network for knowledge graph embedding

Graph Neural Networks (GNNs) are inherently suited for modeling graph-structured data and have been extensively utilized in Knowledge Graph Embedding (KGE). Current GNN-based KGE models primarily focus on message aggregation among entities, often neglecting the aggregation of messages related to relations. Additionally, the interaction information between entities and relations, as well as their distinctions, is overlooked during the updating of relations. To address these issues, we propose the Entity-Relation Aggregation Mechanism Graph Neural Network (ERAGNN), where relations are also considered as nodes in the graph for message aggregation. The ERAGNN layer comprises an entity aggregation sublayer and a relation aggregation sublayer. The entity aggregation sublayer employs an entity-relation composition operation to aggregate messages across entity nodes, while the relation aggregation sublayer utilizes an entity-entity composition operation. Furthermore, shared-weight matrices are implemented to enhance interactions between entities and relations. Lastly, an attention mechanism is incorporated to differentiate neighboring messages during the update of relation embeddings. Experimental results demonstrate that ERAGNN achieves state-of-the-art link prediction performance on three benchmark datasets: FB15k-237, WN18RR, and WN18.