A Poisson Mixture Model for Clustering and Feature Selection of High-Dimensional Count Data

• Activity Number: 601
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Qiwei Li* and Michele Guindani and Brian J. Reich and Howard Bondell and Marina Vannucci
• Companies: Rice University and MD Anderson Cancer Center and North Carolina State University and North Carolina State University and Rice University
• Keywords: Dirichlet process mixture ; Poisson mixture ; Variable selection ; Clustering ; High-dimensional data ; Count data

Abstract

Modern big data analytics often involves high-dimensional count data sets such as bag-of-words data, ecological survey data, and biological sequence data. A main goal in the analysis of such data is to uncover the group structure in the samples as well as identify the discriminating features. We propose a Bayesian nonparametric hierarchical Poisson mixture model that accounts for the overdispersion observed across samples as well as across multiple features. Our model formulation incorporates a feature section mechanism and prior distributions that appropriately account for identifiability constraints on the model parameters. Our strategy for posterior inference results in a unified approach to achieve the goal. We demonstrate the performance of our method on simulated data and present applications to document clustering, based on a bag-of-words benchmark data set, and to document classification, with an analysis of the Federalist Papers.