Research On Improvement Of Algorithm Based On Collaborative Filtering And Generative Adversarial Network And Design Of Recommender System

Today,people have entered the era of information explosion.The research of personalized recommendation system occupies an increasingly important position.In the context of artificial intelligence,deep learning algorithms have achieved rapid development.Among them,the generative adversarial network GAN has made great achievements in the fields of music generation,image generation,sound generation and so on.The Top N recommendation algorithm CFGAN combines collaborative filtering with generative adversarial networks,and solves the problem of discretized indexing through vector adversarial training.The CFGAN recommendation algorithm has achieved good results and has been applied in recommendation systems.However,the real-valued vectors generated by the generator of the CFGAN algorithm still have room for improvement in terms of physical parseability and logical expressibility.In response to the above challenging problems,this thesis proposes a new recommendation algorithm and designs a novel offline recommendation system.The main research work of the thesis is as follows:First,the thesis conducts simulation experiments on the three model methods ZR,PM and ZP proposed in CFGAN to verify the feasibility of the zero regularization penalty term and the partial masking strategy.On this basis,the defects of the real-valued vector generated by the generator of the CFGAN algorithm in terms of physical parseability and logical expressibility are analyzed.Second,the thesis embeds the "1 reconstruction" regularization penalty term for the output nodes corresponding to the positive samples of the real-valued vectors generated by the CFGAN generator.In the real-valued vector generated by the generator,the value of the output node corresponding to the positive sample does not tend to be close to 1 in the same way as the zero regularization penalty is close to 0 as expected.In response to this problem,the thesis embeds the "1 reconstruction" regularization penalty term for the output nodes corresponding to the positive samples.The experimental data show that this innovation can improve the analytical performance of the model and improve the comprehensive performance of the model.Third,the thesis embeds the "1 reconstruction" regularization penalty term for the output nodes corresponding to the partial negative samples of the real-valued vector generated by the CFGAN generator.For the real-valued vector generated by the generator of the CFGAN algorithm,only by virtue of the mutual confrontation between the generator and the discriminator,it is difficult to obtain a score close to 1 for the item finally recommended to the user.To solve this problem,the "1 reconstruction" regularization penalty term is embedded in the output nodes corresponding to some negative samples.Experiments show that this innovation makes the model better.Finally,the thesis names the collaborative filtering generative adversarial network embedded with the double "1 reconstruction" regularization penalty term as PLGAN,and designs an offline personalized recommendation system based on PLGAN.First,experiments are carried out on the public dataset to study the physical parseability,logical expressibility,training speed and accuracy of the PLGAN model.Then an offline movie recommendation system is designed based on the PLGAN algorithm and the big data platform.Finally,use big data technology（such as hadop,spark,kafka,flume,etc.）,background development technology（Spring Boot,Mybatis,etc.）and deep learning algorithm PLGAN to implement the deployment of the recommendation system.In summary,this thesis proposes a new recommendation algorithm.The PLGAN embedded with double "1 reconstruction" increases the physical parseability and logical expressibility of the model,improves the accuracy of the model,and shortens the overall training time of the model.At the same time,based on the PLGAN algorithm,this thesis designs and deploys a novel offline personalized recommendation system that integrates front-end and back-end,which solves the closed-loop problem from data to algorithm and from algorithm to personalized recommendation.