Abstract

In this thesis, we propose and develop various statistical models to enhance and improve the efficiency of statistical modeling of count data in various applications. The major emphasis of the work is focused on developing hierarchical models. Various schemes of hierarchical structures are thus developed and analyzed in this work ranging from purely static hierarchies to dynamic models.
The second part of the work concerns itself with the development of multitopic statistical models. It has been shown that these models provide more realistic modeling characteristics in comparison to mono topic models. We proceed with developing several multitopic models and we analyze their performance against benchmark models. We show that our proposed models in the majority of instances improve the modeling efficiency in comparison to some benchmark models, without drastically increasing the computational demands.
In the last part of the work, we extend our proposed multitopic models to include online learning capability and again we show the relative superiority of our models in comparison to the benchmark models. Various real world applications such as object recognition, scene classification, text classification and action recognition, are used for analyzing the strengths and weaknesses of our proposed models.