| User preference is a latent variable that determines online product ratings and predicting ratings according to user preference is of great interest. This thesis is to construct the latent variable model with user preference, and describe arbitrary dependence relationships as well as the corresponding uncertainties in rating data by adopting Bayesian network as the preliminary framework. In this thesis, we start from the rating data and construct the product rating model without latent variables at first. Then, we give the method for inserting latent variables based the semi-clique structure, so the model can be constructed to describe user preference by the inserted a latent variable. Following, we give the EM-algorithm based method for estimating parameters in the latent variable model. Finally, we propose the algorithm for probabilistic inferences of the latent variable model and the method for predicting user ratings. Experimental results on the MovieLens and Book-Crossing dataset show that our method is effective. The main work of this thesis includes the following three aspects:(1) User preference is represented by a latent variable. We define the latent variable model with user preference called RBNL(Rating Bayesian Network with a Latent variable).K2 algorithm with BIC score is used to construct the product rating model without latent variables at first.(2) We give the method for inserting latent variables based the semi-clique structure, so the model can be constructed to describe user preference by the inserted a latent variable. Following, we give the EM-algorithm based method for estimating parameters in the latent variable model, and use the extended BIC scoring function to select the best hidden variable model.(3) In order to reduce the time complexity of computing the joint probability distribution, we propose the variable elimination algorithm for probabilistic inferences of the latent variable model and the method for predicting user ratings. |
PDF Full Text Request