A Smoothed Latent Generalized Dirichlet Allocation Model in the Collapsed Space

摘要

Many LDA (latent Dirichlet Allocation)-based models often suffer from a large bias due to the variational Bayesian inference scheme with its strong independency assumption between latent variables and parameters. In addition, the quest for better approaches in many instances lead to very sophisticated models (nonparametric). As a result, in this paper, instead of a complex method, we propose a simple framework that only utilizes the collapsed Gibbs sampling inference technique coupled with the flexible GD (generalized Dirichlet) prior to obtain accurate estimations. Experimental results in image categorization show the merits of the new approach.