Personalized Recommendation Based On Heterogeneous Graph Neural Network

With the rapid development of information technology and the Internet,people have entered an era of "information overload".The personalized recommendation is an effective way to solve "information overload",and it plays a pivotal role in life services.In traditional recommendation algorithms,such as collaborative filtering algorithms and matrix decomposition algorithms,usually only consider the historical interaction records of users and items for the recommendation,which leads to deviations in user and item modeling and affects the accuracy of personalized recommendations.In recent years,the heterogeneous graph neural network has been proposed as a modeling method that integrates complex information networks.Since the heterogeneous graph neural network is very flexible in modeling heterogeneous data,it is often used in the recommendation to express rich auxiliary information.The recommendation algorithm based on this setting is called the recommendation algorithm of the heterogeneous graph neural network.In the early days,the recommendation algorithm of heterogeneous graph neural networks was mainly based on the similarity of meta-paths.However,this method could not effectively extract and utilize the rich structural information and semantic information of heterogeneous graphs.Obtaining auxiliary information in a heterogeneous graph mainly presents the following challenges:(1)How to sample heterogeneous neighbors closely related to the embedding of each node;(2)How to design a node encoder to solve the content heterogeneity of different nodes;(3)How to distinguish the subtle differences between node neighbors and select some information-rich neighbors;(4)How to choose the most meaningful meta-path and fuse semantic information for a specific task.In response to the related challenges in heterogeneous graphs,this paper proposes personalized recommendations based on heterogeneous graph neural networks.The specific related work is as follows:(1)A recommendation algorithm based on a heterogeneous feature aggregation deep network,called HFAN,is proposed to fully mine the potential structural features of users and items in heterogeneous graphs.In the data structure,a heterogeneous information graph is constructed and a restart-based random walk strategy is used to sample the heterogeneous neighbors of the node.In representation learning,a hierarchical feature aggregation deep network is designed,including the same type and different types of feature aggregation networks,which can better capture the complex structure and rich semantic information of heterogeneous graphs.In model prediction,the learned final node embedding is transformed by a fusion function and then integrated into the matrix factorization model to complete a specific score prediction.A large number of comparative experiments on three real data sets have verified the effectiveness of the HFAN model in the scoring prediction recommendation task,and the performance is better than the recommendation algorithm based on matrix factorization.(2)A meta-path recommendation algorithm based on a heterogeneous hierarchical attention network,called HHAN,is proposed to model different types of nodes and their rich attributes and interaction relationships.First,the neighbor attribute information of the node is used to enhance the representation of the node,and the space conversion of different types of node embeddings is performed.Then a hierarchical attention network is designed,which can fully consider the node's preference for neighbors and meta-paths,and hierarchically aggregate the characteristics of neighbors based on the meta-path to generate node embeddings,and the learned node embeddings can be better-obtained Structure information and semantic information of heterogeneous graphs,and complete specific Top-K ranking recommendations.A large number of comparative experiments on three real data sets verify the effectiveness of the HHAN model on the Top-K ranking recommendation task,and the performance is better than the recommendation algorithm based on collaborative filtering.