In the era of advanced information and popularization of mobile Internet,people are faced with massive data information,and it is difficult to quickly find the content that meets their personalized needs,resulting in information overload.Recommender systems provide accurate and personalized recommendations by deeply mining user interest,behavior data and other information,so as to alleviate information overload.In recent years,researchers have explored applying deep learning techniques to recommender systems.Among them,Graph Convolutional neural network(GCN)has strong representation ability on graph data structure and attracts attention.With its rich structural information and semantic information,Knowledge Graph(KG)has become a research hotspot to solve the problems of data sparsity and cold start.Therefore,many recommendation systems introduce graph convolutional neural networks and knowledge graphs in order to fully mine user preferences and item attributes for personalized recommendation.However,these systems often only focus on the collaborative information of the user-item interaction graph or the item attributes in the item knowledge graph,and few models can simultaneously capture effective information from the user-item interaction graph and the item knowledge graph.In addition,they also introduce a large number of training parameters to increase the model complexity and training difficulty.In the above context,this paper proposes a knowledge graph-based graph convolutional neural network recommendation model SKGCN and combines with the current frontier research trends to carry out targeted comparative experiments on the existing problems.Firstly,the model introduces two specific aggregation layers,LGC and EUCGC,which act on the feature aggregation process of user-item interaction graph and item knowledge graph respectively,so that entities can obtain rich high-order neighborhood information in a hierarchical iterative manner,which solves the problem of insufficient high-order information in other recommendation models.Second,experiments show that removing the feature transformation matrix and nonlinear activation function in the graph convolutional layer can not only reduce the training parameters but also improve the recommendation performance.At the same time,a new attention mechanism function is designed to focus on the importance of entity relationships for specific users and central nodes.In this way,not only the user’s personalized interest in entity relations can be obtained,but also the semantic information of the relations can be enhanced.Thirdly,the final embedding vectors of users and items are obtained by weighted summation,and then brought into the prediction scoring function to generate recommendations.The SKGCN model utilizes the information in two kinds of source data,which effectively solves the cold start problem when only focusing on user-item interaction and the lack of collaborative information when focusing on item knowledge graph.The experimental results on Movie Lens-20 M,Book-Crossing and Last-FM datasets show that the proposed model is significantly better than other mainstream recommendation methods,and the performance comparison in the cold start scenario also shows the superiority of the model.At the same time,we also conduct ablation experiments to explore the impact of different hyperparameter changes on the performance of the recommendation system in EUCGC. |