Abstract

Increase in use of mobile devices and the introduction of cloud-based services have resulted in the generation of enormous amount of data every day. This calls for the need to group these data appropriately into proper categories. Various clustering techniques have been introduced over the years to learn the patterns in data that might better facilitate the classification process. Finite mixture model is one of the crucial methods used for this task. The basic idea of mixture models is to fit the data at hand to an appropriate distribution. The design of mixture models hence involves finding the appropriate parameters of the distribution and estimating the number of clusters in the data. We use a variational component splitting framework to do this which could simultaneously
learn the parameters of the model and estimate the number of components in the model. The variational algorithm helps to overcome the computational complexity of purely Bayesian approaches and the over fitting problems experienced with Maximum Likelihood approaches guaranteeing convergence. The choice of distribution remains the core concern of mixture models in recent research. The efficiency of Dirichlet family of distributions for this purpose has been proved in latest studies especially for non-Gaussian data. This led us to study the impact of variational component splitting approach on mixture models based on several distributions. Hence, our contribution is the application of variational component splitting approach to design finite mixture models based on inverted Dirichlet, generalized inverted Dirichlet and inverted Beta-Liouville distributions. In addition, we also incorporate a simultaneous feature selection approach for generalized inverted Dirichlet mixture model along with component splitting as another experimental contribution. We evaluate the performance of our models with various real-life applications such as object, scene, texture, speech and video categorization.