Research And Implementation Of Cross-Domain Recommendation Method Based On Heterogeneous Graph

Recommendation methods,as a personalized information filtering tool,can effectively alleviate the information overload in the smart city.However,the traditional recommendation methods can only recommend in a single domain,which cannot meet the demands of cross-department,and cross-businesses service recommendations in the smart city.The cross-domain recommendation methods can assist the target domain in recommending with the help of the information from other domains,and solve the problems of cold start and data sparsity,to effectively meet the diversified demands of users.This thesis carries out research from two aspects:cross-domain bundle recommendation and privacy preserving cross-domain federated recommendation,to realize the effective utilization of information in different domains and the precise matching of demands.Aiming at the problem that the existing cross-domain recommendation methods can only leverage the information from the source domain to unidirectionally recommend on the target domain,and cannot achieve the recommendation of multiple related items,a dual-target cross-domain bundle recommendation(DT-CDBR)model is proposed.Firstly,in each domain,a heterogeneous graph with user,item,and bundle(combination of items)as nodes is constructed to model the interaction and affiliation relationships between them.Secondly,the graph convolution network is applied to handle these relationships to extract the features of users and bundles.Thirdly,the weight of each domain is dynamically adjusted based on the attention mechanism to establish the relationship between the features of common users in two domains.Finally,neural network technology is used to describe the features of users and bundles,to predict the probability of interaction,and then implement the bundle recommendation using two domains information.In this thesis,experiments are conducted on real-world datasets,the experimental results show that compared with the typical recommendation models,the DT-CDBR model can make full use of the information in both domains to achieve bidirectional bundle recommendation,and the recommendation performance of the model is significantly improved.Further,the full fusion of multiple domains information can achieve more effective recommendations.Aiming at the problem that the existing cross-domain recommendation methods cannot make full utilization of multiple domains information,and the transfer of information between domains also increases the risk of user privacy leakage,a privacy preserving multi-target cross-domain recommendation(PP-MTCDR)model is proposed.The model applies federated learning technology to collaborative train among multiple local domains and the server.In the local training stage,a heterogeneous graph is constructed to model the interaction between users and items,and the graph convolution network is used to extract the features of users and items from the graph.To prevent user privacy from the model gradient,Rényi-differential privacy technology is applied to add noise to the model gradient to achieve user privacy preserving with accurate noise control.In the server aggregation stage,the local model parameters of privacy preserving in multiple domains are fused to implement the effective utilization of multiple domain information.Each domain is executed alternately with the server,and collaborative trains a cross-domain recommendation model that can protect privacy to meet the effective recommendation of multiple domain fusion.Theoretical analysis proves that the PP-MTCDR model meets(κ,ε)-RDP,which can protect user privacy in the process of multiple domain data fusion.The experimental results show that the PP-MTCDR model can make full use of the information in multiple domains for recommendation and improve the recommendation performance in multiple domains simultaneously.