A Personalized Recommendation Method Based On Heterogeneous Graph

Personalized recommendation can help users filter useless information and quickly locate the content that users are interested in,which greatly improves user experience and has become an indispensable tool in online services.The recommendation algorithm determines the performance of the recommendation system.Among various recommendation algorithms that integrate different cutting-edge technologies,owing to the advent of the era of big data and the improvement of computing power,the recommendation algorithm based on deep learning has been greatly improved performance of the recommendation model.Among them,the recently popular Graph Neural Network(GNN)is widely used in all walks of life because it can easily aggregate the features of nodes and learn the topology around nodes well.However,the existing GNN-based methods often focus on the modeling of user-item bipartite graphs,and adopt a relatively simple fusion method for the aggregation of auxiliary information of nodes,which cannot mine high-order information in hidden node features and damages the generalization ability of the model.In view of the above problems,this thesis improves the traditional GNN model.Based on the original user-item bipartite graph,the corresponding feature graph and knowledge graph are constructed.In order to enrich the final representation of users and items,improve the performance of the model and achieve the purpose of improving the accuracy of recommendation.The main research contents of this thesis are as follows:(1)Feature Graph Collaborative Filtering for Personalized Recommendation Model(FGCF).FGCF uses the feature map of users and items based on auxiliary information to deeply mine users’ preferences to improve the recommendation performance.FGCF uses graph neural network to mine the latent cooperative information and high-order feature information in these two graphs respectively,and then calculate the final prediction score.The experimental results show that the proposed model has a certain degree of improvement over the traditional GNN model based on bipartite graph.(2)Feature Interactive Graph Neural Network for KG-based Personalized Recommendation(FIKGRec)introduces the knowledge graph of items on the basis of FGCF to develop the performance of recommender system.The existing graph neural network methods that integrate knowledge graph information ignore the node interaction in the process of information propagation and information aggregation,which will greatly limit the performance of the model.The overall framework consists of three components:For the item-side modeling,FIKGRec designs an improved graph neural network algorithm to mine the latent structural information and semantic information in the knowledge graph,and combine the collaborative signal in the user-item bipartite graph and the knowledge signal in the knowledge graph.The knowledge signals are organically combined,and get the final representation of items.For the user-side modeling,a preference-aware attention mechanism is designed to obtain the user’s fine-grained preference for items that have been interacted.The experimental results show that the proposed model not only greatly improves the performance of the model,but also greatly alleviates the data sparsity and cold-start problems.