Mixed Recommendation Algorithm Based On Topic Models

With the development of the Internet, the growth of data on the network has been very rapid now. How to find the information users need from massive data has become increasingly important. In era of social networking web2.0 became very popular, social networks text and traditional text has a very different, and how to model these text information, and make recommendations to the user has become a more popular field of study.Topic model can mine the semantic information of the text. The model is established on thematic level, which is the Feature that traditional text model does not have. In the era of social networking, topic model can overcome the problem : text with small numbers of words and big noise. In this paper, I implement two improved LDA models based on the traditional LDA model: ATM-LDA and MB-LDA. ATM-LDA is modeling on author-topic level instead of document-topic level to mine the author’s theme. MB-LDA introduced the relationship of users and documents while modeling the texts, so it can distinguish the topic of different users in one text.After the analysis of current recommendation technology and topic model,a new user interest model based on user interest character and key words sequence is proposed. This model is on the basis of Author-Topic model and key words of users. Latent topic is introduced into topic model and we can compute the the latent interest of user by Gibbs Sampling.Via the combination of user interest character and key word sequence,we can build the tow dimensions of user interest model: latent semantic and content. On the basis of user interest model,a hybrid recommendation algorithm is propossed. This hybrid algorithm combined author-topic model and vector space model, we can compute the user interest distribution through author-topic model and modeling the user key word sequence through vector spacemodel. The tow different similarities is combined linearly to make a recommendation. The experimental data is collected from sina microblog,the results indicated that this hybrid recommendation alorithm has achieved valid effect on precision an diversity.