Research On Personalized Recommender System Based On Deep Transfer Learning

With the rapid development of Internet technology,more and more terminal equipment can quickly and conveniently access the Internet,send and receive information on the Internet.The access to terminal equipment has also enriched the information contained on the Internet,but various kinds of information are woven together and are not easy for people to search.The stored information is too much,which causes the problem of information overload.Under the background of information overload,the recommender system emerged at the historic moment and became an effective way to solve this problem.The recommender system analyzes user's behavior,constructs a unique user portrait for each user,predicts the user's preferences and recommends related content.The recommender system is based on analyzing the historical behavior information of users.Although a lot of information on the Internet,the data that can be processed and used is generally scarce.Therefore,recommender system faces the problem of sparse data in many cases,resulting in lower quality of recommendations for users.Among many strategies for solving data sparseness,it is widely used to combine multiple algorithms to mix recommendations and merge other data sources for recommendation.Through deep learning methods,multi-faceted attribute information is merged to mine the association between user feature matrix and item feature matrix to improve the quality of recommendations.The deep learning model uses a back-propagation algorithm for self-iterative optimization to reduce errors.However,the performance of deep neural network model depends largely on the amount of data.Sufficient data can better train the model,which has prompted research in the field of transfer learning.Transfer learning technology applies the features or knowledge learned in a domain to other related but different domains to improve the performance of model in that domain.By learning and mining a large amount of scoring data,refining the hidden pattern migration and applying it to target domain,initializing the scoring model of target domain to a certain extent,alleviating the data sparseness problem in target domain.The following researches were carried out in this paper:1)A service quality prediction model based on matrix decomposition of transfer neural network was proposed.Through the study of neural network transfer,the information in source domain is used to assist the prediction of the neural network in target domain.Use the information of service node throughput,assist the response time prediction in target domain and recommend service nodes with low response delay for users.The validity of the migration was verified by choosing different data sparsity in the experiment.2)A transfer recommender model based on user-item cross attention was proposed.Based on the study of neural networks,combined with the attention mechanism to fully mine the scoring model of dense data in source domain and transfer it to neural network model in target domain.Use the attention mechanism to find important information on a global scale and give higher weight to critical information.In order to explore the validity of the algorithm model,different data were selected for verification in the experiment.The comparison of the evaluation indicators verified the effectiveness of the model.3)Research the application of recommender system based on Django framework.Based on the existing recommender model,a working recommender system was built by using the open source web framework Django.Learned the MVT idea,applied the recommender model to the project and the problem in project development was solved.