Discriminative LDA
Weiran Xu, Mingzhi Dong, YunHang Lin, Jun Hai Guo, Guang Chen · 2010 2nd IEEE InternationalConference on Network Infrastructure and Digital Content · 2010
This paper is aim to improve the discrimination capability of LDA model through unsupervised feature selection. Experimental results show that if the interference of general word and general topic can be removed, the discrimination capability of LDA model will be increased. The key problem is how to find supervised information to evaluate features. The LDA topics are assumed reasonable. Therefore, topics will offer surprised information for word features' selection. Constraint coming from the surprised information is added to the LDA objective function. Finally, a heuristic algorithm is presented to obtain the solution. Experiments show that the Discriminative LDA can significantly improve the information gain of topics.