Mixture model based data clustering

The study focused on exploring the use of mixture models for data clustering techniques on data with complex heterogeneous properties. Both parametric and nonparametric mixture models were here considered. In particular, Poisson mixture regression models and Zero Inflated Poisson mixture models were applied and analysed for the parametric case as mixtures of Gaussian kernel densities were considered for the nonparametric case. Although, parametric mixtures assume clusters and components to be the same, these are observed different by either using graph theory techniques via voronoi diagrams or infinite componential models in the nonparametric case. Determination of better clustering method using mixture models was enabled through classification via clustering analysis. Results shows that although number of clusters from the parametric mixtures can be the same with nonparametric, these approaches are generally very different. Nonparametric mixtures produce better clusters as evidenced by their ability to improve classifier performance and hence compacting heterogeneity in a more efficiently way whilst parametric are a bit rigid. The ability by mixture models to allow inferences on cluster properties via densities and parameters over traditional clustering methods makes them a better clustering methods. We conclude that although mixture models are better than hierarchical models, nonparametric mixture should be preferred over parametric as they are more robust, flexible and produce better clusters.