| Topic model is a statistical method used to find the hidden theme in the article, which is a very wide range of application. The model assumes that each document exists many hidden topics, and because of the difference of the nature and type of every document, making the topic distribution of each article is different. Topic mod-el is the use of this feature of article and builds model based on the topic distribution of each article. The application scope of the topic model is very wide, and the application effect is very significant in many fields. This article is to the LDA (Latent Dirichlet Analysis) topic model for the study, improvement and estimate model parameters for the model were further studied.First of all, taking into account the e-commerce, we can score goods after purchased them. Therefore, we have improved the LDA model, adding to the score of the goods. In addition, we improved the LDA model according to the discrete model and the continuous score, and then used the variational EM algorithm to estimate the parameters.Secondly, because in the process of the parameters estimation of the improved LDA model by using the variational EM algorithm, we find that the lower bound of the objective function is the entropy of the variational distrib-ution. The entropy is the uncertainty of the random variables, that is, we ass-ume that the uncertainty of the variational distribution. Therefore, we propos-e to use TEM (Expectation Maximization Tempered) algorithm to estimate the parameters of the model. Through the entropy of the objective function divid-ed by a parameter Ï„, thereby reducing the impact to the model due to the un-certainty of the assumption of the variational distribution.Finally, we apply the model to the MovieLens data set, and compare it with the item based collaborative filtering, the user based collaborative filteri-ng and the PLSA model by the method of cross validation. And the improved model is better than the other models, but for the two kinds of model, contin-uous data model is better than the discrete model. |
