Supervised topic models for multi-label classification

Recently, some publications indicated that the generative modeling approaches, i.e., topic models, achieved appreciated performance on multi-label classification, especially for skewed data sets. In this paper, we develop two supervised topic models for multi-label classification problems. The two models, i.e., Frequency-LDA (FLDA) and Dependency-Frequency-LDA (DFLDA), extend Latent Dirichlet Allocation (LDA) via two observations, i.e., the frequencies of the labels and the dependencies among different labels. We train the models by the Gibbs sampler algorithm. The experiment results on well known collections demonstrate that our two models outperform the state-of-the-art approaches.