Personalized News Recommendation System Based On Topic Modeling And Hierarchical Latent Variable Models

News has the characteristics of short life cycle,sparse records,and the complexity of the text.Because of the characteristics of the news,we construct a news recommendation system based on topic modeling and hierarchical latent variable models in the paper.The recommendation system uses the hybrid recommender model,which combines the content-based recommendation algorithm and the collaborative filtering recommendation algorithm,to recommend news for users.The model we construct in the paper includes three modules:(1)news topic modeling,(2)news clustering based on topic features,(3)hierarchical latent variable models.The hybrid recommender model uses LDA algorithm to model the topic features of the news.After news modeling,we can obtain the distribution of topics and words and the probability matrix of the topics.We use the self-organizing neural network(SOM)to cluster the news according to the probability matrix of the topics.Self-organizing neural network has better visualization effects.We can effectively determine the number of news clusters based on the clustering result.In the news recommendation system,we make use of the user's access log to structure the users' pseudo rating matrix for the news.In addition,the model divides the users' scarcity rating matrix into two low-dimensional non-sparse matrices.It uses the users' implicit features,the topic features of news,and two low-dimensional matrices to approximate the original rating matrix.The difference between the model constructed in the paper and the traditional latent variable models is that our model linearly integrates the topic features of the news and the users' implicit feature into the latent variable model.We import the idea of threshold autoregressive,and divide the linear fusion model into a number of linear fusion models.Moreover,in order to improve generalization ability of the model,we add different kinds of regularization items in the model.In this paper,we use the root-mean-square error as the evaluation index.We evaluate the performance of the hybrid recommendation algorithm proposed in the paper,which based on topic modeling and hierarchical latent variable models.And we contrast the performance of the proposed algorithm with other matrix decomposition algorithm and the collaborative filtering recommendation algorithm.We found that the hierarchical latent variable models based on the topic features of news can reduce the root mean square error about 10%-20%,compared with other algorithms in the data set of Xiamen University news reading website.In the paper,we also introduce the process of parameter selection.We choose the suitable topic number of the news,the clustering category of the news and the optimal parameters of the hierarchical latent variable models for the data set of Xiamen University news reading website.