Transelliptical Topic Modeling with Application to Genomics Data

• Activity Number: 614
• Type: Contributed
• Date/Time: Thursday, August 7, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Xingyuan Fang*+ and Han Liu
• Companies: Princeton University and Princeton University
• Keywords: High-dimensionality ; Big Data ; Microarray Data

Abstract

We propose a new topic model for exploring and analyzing high dimensional nonGaussian data. The core of our approach is transelliptical modeling, which provides a flexible semiparametric framework for modeling complex heavy-tailed data. Our approach has several important features: (1) Unlike many existing topic models which are constructed in a parametric way to reflect the generating mechanism of documents, our approach is non-generative and semiparametric; (2) Our model allows the estimated topic profiles to contain negative entries, thus is more realistic for modeling microarray data; (3) Our approach is suitable for modeling heavy-tailed data with nontrivial tail dependence. Computationally, our method is efficient and scales to very large datasets. Theoretically, the estimated topics attain parametric rates of convergence. To illustrate the usefulness of the proposed method, we apply it on publicly available gene expression data and ChIP-chip or ChIP-seq data to conduct high-throughput association analysis between a large number of transcription factors and biological contexts. Our results are supported by existing biology literatures. For example, we find that a human transcr