Research On Recommendation Algorithms Based On Knowledge Awareness And Graph Neural Network

With the rapid development of computer technology and the emergence of 5G technology,digital information has emerged in all aspects of our lives,and the magnitude of data has also increased exponentially.In the face of all kinds of complex data,people are often helpless and difficult to mine the hidden link between data.The emergence of the Recommendation System has greatly alleviated the pressure of data overload and provided users with more effective purchasing reference.However,most traditional recommendation methods only establish explicit and implicit feedback relationships,without taking into account the higher-order relationship between itemitem,user-user and user-item.This leads to serious data sparsity problems.In recent years,in response to the above problems,researchers have used the knowledge map as the auxiliary input of the recommendation system to establish multiple connections between users and items.They use the graph neural network propagation algorithm to further mine the embedding information of user and item nodes.By incorporating the user’s social information,user’s multibehavior information and the association information between items into the model,the accuracy of the recommendation system will be improved and the user will have a better shopping experience.Although the improvement of these methods has achieved certain results,there are still many shortcomings.In this thesis,based on the existing problems in the current model,we have studied the time-series characteristics of user-item interaction,user multi-behavior characteristics,user social information and the impact of graph network propagation algorithm.The research contents of this thesis are as follows:(1)In view of the lack of analysis of user multi-behavior feature interaction information in the recommendation model,a new hybrid graph neural network model based on user multibehavior interaction(UMBGN)is proposed.The model uses the joint learning mechanism to perceive the sequence information of user-item multi-behavior interaction.This model focuses on the long-term multi-behavior characteristics of users according to the user multi-behavior information perception layer,and uses Bi GRU unit and AUGRU unit to learn time-ordered useritem interaction information.In addition,our model also defines the propagation weights between the user-item interaction diagram and the item-item relationship diagram according to the user behavior preferences to obtain more valuable dependencies.(2)The traditional method only considers the user’s personal characteristics when mining user preferences,and does not consider that the user’s social information will affect the user’s decision and shopping intention.In order to further mine user preference information,a new hybrid graph neural network model(UPSGN)based on user preference characteristics and social information is proposed in this thesis.The model uses graph network to extract user social information and embed it through user social information extraction module.Then,we use the attention mechanism to get the user embedding representation from three aspects: user social information,user multi-behavior characteristics,and user item interaction characteristics,which solves the single problem of only considering user behavior characteristics.(3)We have carried out a lot of experiments on the proposed UMBGN model and UPSGN model on five real data sets,and made a quantitative and qualitative analysis of the role of each module of the model.The experimental results show that the two models proposed in this thesis have significantly improved compared with the contrast model,which proves the effectiveness of the model.