User Preference Modeling Based On Deep Belief Network And Latent Variable Model

With the rapid development of the mobile Internet,the Internet has entered all aspects of our lives,but also produced a large number of user behavior data.These user behaviordata have the characteristics of massive,high-dimensional,complex internal structure,and contain user preferences.User preferences express the user's personal preferences and possible behavior.Therefore,it is of great significance to build user preference model,score prediction and preference estimation based on preference model,and provide effective support for personalized service and recommendation.On the one hand,user preferences exist objectively but can not be observed directly.Implicit variables are used to describe variables that cannot be observed directly.At the same time,there are complex interdependencies among attributes in user rating data,and Bayesian Network(BN)can effectively express arbitrary dependencies and uncertainties among attributes,and it has good reasoning ability.Introducing latent variables into Bayesian network is an effective method to construct user preference model?which has been widely used in the field of uncertain knowledge.Therefore,this thesis considers that implicit variables are used to represent user preferences,and a user preference model based on implicit variable model in scoring data is proposed.On the other hand,the process of constructing Bayesian networks with implicit variables will produce a large number of intermediate data,which makes the computational complexity increase sharply.Therefore,this thesis considers introducing Deep Belief Network(DBN)classifier to construct user preference model on the basis of Bayesian networks with implicit variables to reduce the complexity of model construction.Specifically,this thesis uses DBN to classify the scoring data,and expands the implicit variable model with category variables,which is called Class Simplified BN(CSBN).Then,based on the characteristics of scoring data and the key steps of implicit variable model construction,the constraints to be satisfied in model construction,as well as the parameters learning and structure of model under constraints are given.Learning methods.In addition,scoring prediction and preference estimation are two important applications of user preference model.They are the direct support of personalized services and recommendation,and can be accomplished based on Bayesian network probabilistic reasoning.Therefore,based on the basic framework of CSBN model and variable elimination method,this thesis presents the method of preference estimation and score prediction of scoring data.Finally,based on the MovieLens dataset and the public comment dataset,the method proposed in this paper is tested to verify the efficiency and effectiveness of this method.