Dynamic threshold model based probabilistic latent semantic analysis

Abstract

Probabilistic Latent Semantic Analysis(PLSA) is the one of the main methods for texture analysis and computer vision. In practice, PLSA will result in overfitting problems, including the circumstance of unclear membership of topics and the case of high similarity between different topics. In this paper, we describe a dynamic threshold model based PLSA(dPLSA). It can make the ambiguous topic information more clear and objectified. Meanwhile, dPLSA can dynamically determine whether to merge the similar topics, in terms of the potential similarity between different topics. Experimental results on image data sets show that the proposed method outperforms its rival ones for solving the overfitting problems.