A Novel Term Selection Approach in sLDA for Imbalanced Text Categorization
Zhenyan Liu, Dan Meng, Weiping Wang, Yong Wang, Chenhao Bai · Advances in computer science research · 2015
The supervised Latent Dirichlet Allocation (sLDA) is a probabilistic topic model of labelled documents, which is better than unsupervised LDA for text categorization.But sLDA experiments were based upon this default assumtion that the corpus is balanced, that is, the samples of each class are approximately equal, and chose a vocabulary by tf-idf.While the corpus is imbalanced, tf-idf tends to choose terms from the majority classes and ignore terms of the minority ones.Thus the performance of text classifier will be degraded severely.Therefore this paper proposed a new term selection approach which can fairly choose more discriminative terms from every category.Experimental results show that using this new approach in sLDA for imbalanced text categorization can greatly impove the recall and precision of the minority classes, and it is superior to tf-idf.