Knowledge Graph Recommendation Based On Graph Attention Network

Data sparsity and cold start issues have always been difficult problems faced by traditional recommendation algorithms.In recent years,more and more research has considered incorporating some auxiliary information into the recommendation process,such as knowledge graph,user social networks,etc.,to enhance the accuracy,diversity,and interpretability of recommendation results.The combination of these auxiliary information and the user item interaction bipartite graph forms a graph structure that contains rich entity information and structural information.Graph neural networks have excellent performance in processing graph structure data,and can fully mine node information and topology information in graph structures.Currently,more and more knowledge graph recommendation and social networks recommendation based on graph neural network models for recommendation prediction.However,the existing knowledge graph recommendation methods based on graph neural networks either randomly initialize the input vectors,causing them to lose the structural information in the knowledge graph;Alternatively,when processing node information in the graph,they should be treated equally and processed in an averaging manner,which does not reflect the different influences between nodes.In response to the above existing problems,this thesis has conducted the following research:(1)A knowledge graph recommendation model(TDGAT)combining Trans D and graph attention networks is proposed.The TDGAT model first constructs a mapping vector for each pair of entities and relationships in the knowledge graph by using the Trans D model in the KGE method at the embedding layer,obtaining an embedded representation that retains semantic information and structural information of the knowledge graph.After using it as input,it aggregate updated neighborhood information in the graph attention network,and uses a multiple attention mechanism to prevent model overfitting,Finally,the embedded representation obtained at the embedding layer and the embedded representation obtained at the graph attention layer are spliced as vector representations of the final user and item for recommendation prediction.The entity embedding representation obtained by the Trans D model used in the embedding layer of the TDGAT model not only contains more information than other KGE methods,but also greatly reduces the amount of parameters used,reducing the complexity of the model;In the graph attention layer,attention mechanisms are used to automatically assign weights based on the influence of users and items.Experiments on data sets such as Movie Lens and Amazon Book have also proven that the proposed TDGAT model performs better than the baseline model.(2)Aiming at the current problems of social network recommendation that do not take into account user preferences and fail to fully explore the high-level transfer information between users and items,a social network recommendation model(SNRS-GAM)incorporating graph attention mechanism is proposed.By using the graph attention mechanism to model and learn on bipartite graphs and social network graphs,we can obtain user embedded representations and item embedded representations from different perspectives,as well as mine high-level transfer relationships between users and items.Comparative experiments and parameter sensitivity experiments on Ciao and Epions datasets also demonstrate the effectiveness of the proposed model.