Construction And Inferences Of Latent Variable Model Oriented To User Preference Discovery From Sparse Rating Data

Currently,discovering user preference is an important component of Intelligent Internet,which can be utilized by service provider to push individuation information to users.Therefore,people's pay more attention to the analysis of the users' behavior on the Internet and exploring its commercial value.User rating data about products generated in e-commerce may express users' views and preferences.But the data always were sparse on account of some users with little or none score behavior.So how to discover user preference accurately in sparse user rating data is an important foundation of individual service and accurate marketing.Bayesian Network(BN)is a knowledge framework for uncertainty representation,inference and analysis.Though user preference is objective existence,we can not directly observe the value.We can use the Latent Variable in BN to express user preference.Since BN can only describe the uncertainty dependencies between observable attributes,the Latent Variable is missing,so we can not use the Maximum Likelihood Estimation(MLE)method to calculate the Conditional Probability Table(CPT),which can not inference based on BN.For this reason,we introduce the Latent Variable Model to describe the uncertainty between attributes of hidden variables and use the Expectation Maximization(EM)algorithm to calculate the CPT of Latent Variable.To solve the above problems,our work covers the following three contents:(1)Building Latent Variable model based on sparse user rating data.At first,we use the Biased Matrix Factorization model to fill up the scarce rating data,which can help to get the integrated commodity rating data.Then,the commodity BN model based on Mutual Information will be shown.Finally,the way of building the latent variables model based on Max Semi-clique and Expectation Maximization will be given,which can help to get CBN with Latent variable model(CBNL).(2)CBNL probabilistic reasoning oriented user preference discovery.As for the CBNL model,we show the approximate probabilistic reasoning of CBNL on account of Gibbs sampling algorithm,which can calculate the uncertainty of latent variable evaluation by given evidence value.Then we can discover the user preference efficiently.(3)Experimental test.We use the MovieLens dataset to test the correctivity and effectiveness of approach in this thesis.The experimental results showed that the Gibbs sampling algorithm based on Latent variable model to discover the user preference is viable in some degree.