Research On Cross-domain Recommendation Algorithm Based On Machine Learning

Cross-domain recommendation is an important recommendation method,which utilizes information from rich domains to solve the data sparsity and cold-start problems in sparse domains.Cross-domain recommendation needs to rely on high-quality user and item features,while most existing cross-domain methods extract features from user-item rating matrix,which has the problem of insufficient feature quality.In addition,existing cross-domain models often train a general mapping network for common users between two domains,ignoring the differences between users.Ideally,each user should have a personalized mapping model.Insufficient feature quality and lack of personalization both negatively impact recommendation performance.With the increasing importance of privacy security,recommender systems need to consider the issue of user privacy protection more.As an emerging distributed machine learning framework,federated learning can solve the problem of user privacy protection,but it also faces the problems of insufficient privacy protection and high communication load while protecting privacy.In recent years,the development process of cross-domain recommendation has presented many challenges.This paper proposes optimizations on the recommendation performance and privacy protection of cross-domain recommendation.The work is summarized as follows:First,for the problem of feature quality,a cross-domain recommendation algorithm GECDR based on graph embedding is proposed.GECDR first constructs a heterogeneous graph containing users and items,then uses graph embedding technology to extract accurate user and item features from the graph,and then applies two optimization strategies,distance-oriented and task-oriented,respectively,when training cross-domain models.Multiple sets of experiments are conducted on a cross-domain task.The experimental results show that compared with the classical cross-domain model EMCDR,GECDR can reduce the root mean square error(RMSE)of the recommended index by 25.14% and the mean absolute error(MAE)by 18.82%.Second,for the personalized cross-domain problem,a meta-learning-based personalized cross-domain recommendation algorithm Meta＿GECDR is proposed.On the basis of GECDR,to further address the deficiencies of existing cross-domain methods,a metalearner is proposed to generate a private mapping model for each user.We analyze the influence of hidden feature dimension and optimization strategy on recommendation performance and convergence speed in detail in the experimental section.The experimental results of recommendation performance show that Meta＿GECDR can further improve the recommendation quality compared with GECDR,the best RMSE can be reduced by5.46%,and the best MAE can be reduced by 3.03%.Third,for privacy protection issues,a cross-domain recommendation algorithm FedCDR based on federated learning is proposed.Fed CDR follows the basic paradigm of federated learning,saving user-related data on the device side,and mobile devices perform distributed training by downloading the latest parameters and uploading gradients to the cloud center server.We propose a data protection strategy to transform the source domain features and share them with the target domain,and propose a user selection strategy to reduce the noise impact of low-participation users.The experimental results show that compared with EMCDR,Fed CDR can best reduce RMSE by 5.35%,MAE can reduce by 3.55%,and the training time can increase by up to 32%.The additional training time brought by federated learning is within a reasonable range.Fourth,for the problem of communication load,we propose a lightweight federated cross-domain recommendation algorithm Meta EM.When the number of items is too large,the download of the item matrix and the upload of the gradient will cause serious communication overhead to mobile devices.To address this issue,we significantly reduce the traffic on mobile devices by leveraging a meta-learner to generate private and smallerscale item feature matrices for users.The experimental results show that Meta EM has better recommendation performance than the best benchmark methods,the best RMSE can be reduced by 34.22%,the best MAE can be reduced by 32.85%,and it is more compatible with complex embedding models.