Modeling And Application Of Multi-Dimensional User Preference Based On The Latent Variable Model

With the rapid development of mobile Internet and Web2.0,the Internet has penetrated into every aspect of people's life and work,and a large number of user's behavior data has been generated.User's behavior data contains user's preferences,and user's preferences indicate the user's personal tendencies and possible behaviors.At the same time,users tend to have multiple aspects of rating objects.For example,user's tendency towards a movie can be divided into many aspects,such as type,language and so on,which forms the multiple dimensions of the user's preferences.Therefore,understanding the user's behavior data effectively,building multi-dimensional user's preference models,predicting user's ratings based on preference models and estimating user's preferences are important support for personalized services such as precise recommendation and behavior orientation.User preferences are objective,but cannot be observed directly.Latent variables can describe variables that cannot be observed directly.At the same time,the dependencies between the attributes of the user's rating data are arbitrary and uncertain,besides,the Bayesian network can effectively express any dependencies between attributes and uncertainties,and it has good reasoning ability.The introduction of latent variables to Bayesian network is an effective method for user's preference modeling,which has been widely used in uncertain knowledge domain.In this thesis,a multidimensional user preference model is built with a latent variable model with multiple latent variables.In the case of multiple latent variables,a large number of intermediate data will be generated during the process of modeling,and the computational complexity will increase dramatically.The Spark computing framework can effectively handle the problem of high computational complexity.Preference estimation and rating prediction are two important applications of multidimensional user preference models,they are the direct support of personalization and other services,and the core of them is Bayesian network reasoning.Traditional reasoning methods have more repeated calculation and higher computational complexity.Variable elimination method is an effective Bayesian network inference method to reduce repeated calculation.The main work of this thesis is summarized as follows:1.In this thesis,multi-dimensional preferences from rating data are described by multiple latent variables and the Bayesian network with multiple latent variables is adopted as the preliminary knowledge framework of user preference.2.Constraint conditions are given according to the inherence of user preference and latent variables,upon which we propose a method for modeling user preference.Parameters are computed by EM algorithm and structure is established by SEM algorithm with respect to the given constraints,and implemented by Spark computing framework.3.We propose a method of preference estimation and rating prediction algorithm based on a method called variable elimination with the idea of preprocessing,and implemented by Spark computing framework.Experimental results on the Movielens dataset verify that the method proposed in this paper is effective.