Infinite Von Mises-Fisher Mixture Model and Its Application to Gene Expression Data Clustering

In some applications of data mining, clustering analysis of directional data is often involved. In this case, conventional model-based clustering methods are not suitable for fitting the data of such type of structure. Therefore, a Dirichlet Process Mixture Model based on von Mises-Fisher distribution was proposed for the clustering analysis of directional data. The main motivation is that as a non-parametric Bayesian model, The Dirichlet process can automatically determine the complexity of the mixture model when the number of data categories is unknown. We use the accelerated Variational inference algorithm to quickly estimate the parameters involved in the model, which enables the method to be applied in applications with large data scale. The validity of the proposed model was verified by using different scale simulation data and clustering analysis of gene expression data.