Variational learning of finite shifted scaled Dirichlet mixture models

Abstract

Abstract In this chapter, we propose a variational framework to learn about the finite shifted-scaled Dirichlet mixture model and evaluate its performance as a clustering tool in three medical real-world applications, namely, malaria detection, breast cancer diagnosis and cardiovascular disease detection as well as a text application, namely spam detection. The motivation behind applying variational inference is that, compared to the conventional Bayesian approach, it is much less computationally costly. Moreover, in this method the optimal number of components is estimated along with the parameter approximation simultaneously as part of the Bayesian inference method. Besides, in the variational technique, over-fitting as one of the common issues of deterministic approaches is avoided. To show the merits of our model performance, it is compared with four other widely used methods.