Research On User Preference Recommendation Technology Integrating Knowledge Graph

With the rapid development of modern information technology,various types of data and information in the real world show a spurt of growth.The recommendation system has become one of the research hotspots in academia and industry due to its powerful information filtering capabilities.Because the knowledge graph is rich in high-quality semantic knowledge,it has been widely used in recommendation systems in recent years.However,when using graph convolutional network to extract the feature information of entities in the graph,the existing model has the following problems: spatial domain graph convolutional network ignores the neighbor relationship between entities,and the neighborhood set obtained by sampling cannot be learned.To high-quality entity features;when the number of convolutional layers in the spectral domain graph convolutional network is too deep,the features of the nodes in the same connected component in the graph tend to be consistent.In response to the above problems,this paper respectively proposes a recommendation model KGCN-PN based on the common neighbor ranking sampling of the knowledge graph and a lightweight graph convolutional network recommendation model KGLGC based on the fusion of the knowledge graph.The main work and innovations of this paper include:1.Aiming at the problem that the spatial graph convolutional network does not consider the neighbor relationship when sampling the neighborhood set,a recommendation model based on the common neighbor ordering sampling of the knowledge graph is proposed.The model first compares each of the knowledge graphs according to the number of common neighbors.The neighborhood of the entity is sorted and sampled;then the entity’s own information and the receiving domain information are merged layer by layer along the relationship path in the graph.In this process,the "user-relationship scoring function" is introduced,which is used to smooth the entity characteristics of the receiving domain.By calculating the product of the feature vector of each entity and the corresponding relationship vector to measure the importance of the entity to the central entity,thereby assigning different weights to it in the feature fusion stage,thereby improving the entity obtained by each layer of neighborhood aggregation Feature quality;Finally,the user feature vector and the entity feature vector obtained by the fusion are sent to the prediction function to predict the probability of the user interacting with the entity item.2.Aiming at the over-smoothing problem caused by the deepening of the number of convolutional layers in the spectral domain graph convolutional network,a lightweight graph convolutional network recommendation model fused with knowledge graphs is proposed.The model first interacts the knowledge graphs with the history of user items The records are integrated into the same dimensional space,and then the graph convolutional network with the feature transformation and nonlinear activation function deleted is used to learn the feature vectors of the user node and the entity node in different layers through linear diffusion,and finally the inter-layer combination is adopted.In the method,the features obtained from each layer are weighted and combined,and the interaction probability of the two is calculated through the vector inner product,and the ranking is recommended to the user.3.Based on the Tensor Flow framework,the KGCN-PN and KGLGC algorithm models were constructed respectively,and comparative experiments were completed with the relevant baseline models on the Movie Lens-20 M and Last.FM data sets.The experimental results show that the recall rate and cumulative gain of KGCN-PN in the cold-start recommendation task with sparse data have been correspondingly improved,and KGLGC can effectively alleviate the excessive smoothing problem caused by the excessively deep graph convolution layer.