Knowledge graph embedding model based on multi-hop adaptive graph attention network

Knowledge graphs use graphs to represent multi-relational data. Graph attention networks have recently demonstrated remarkable efficacy in predicting connections within knowledge graphs. The focus of graph attention networks is restricted to the embedded information from nodes that are first-order neighbors. To obtain multi-order neighbor information, existing works usually use superimposed multi-layer graph attention networks to achieve the purpose of aggregating multi-order neighbor information. However, this methodology could lead to a homogenization of node features in the knowledge graph, thereby posing challenges to the differentiation of features among nodes. We introduce an approach that employs a multi-hop graph attention network to integrate relationships between nodes across multiple hops and entity details, viewing the combined features of nodes from different orders as temporal sequences. The gating mechanism inherent in Long Short-Term Memory (LSTM) networks is utilized to selectively distill and refine the feature information of nodes. The effectiveness of our model is substantiated through evaluations on popular benchmark datasets, with the experimental findings revealing its superior predictive accuracy in knowledge graph link prediction compared to extant models.