A Selective Supervised Latent Beta-Liouville Allocation for Document Classification

We propose a novel model, selective supervised Latent Beta-Liouville (ssLBLA), that improves the performance and generative process of supervised probabilistic topic models with a more flexible prior and simple framework. ssLBLA model utilizes the “bag-of-selective-words” instead of the “bag-of-words” in the topic modeling by using a Bernoulli distribution to identify the discrimination power of a word for its assigned topic. Indeed, ssLBLA improves and inherits the general framework of selective supervised Latent Dirichlet Allocation (ssLDA) and can predict many types of responses. This paper presents a simple framework that utilizes the collapsed Gibbs sampling inference technique coupled with the flexible Beta-Liouville (BL) distribution prior to achieve more accurate estimations. Experimental results in single-label document classification show the merits of our new approach.