Bayesian Latent Class Models

Xu, Jianfeng

Permalink

https://hdl.handle.net/2142/24418 Copy

Description

Title

Bayesian Latent Class Models

Author(s)

Xu, Jianfeng

Issue Date

2011-05-25T14:36:14Z

Director of Research (if dissertation) or Advisor (if thesis)

Liang, Feng

Doctoral Committee Chair(s)

Liang, Feng

Committee Member(s)

• Simpson, Douglas G.
• Marden, John I.
• Ji, Yuan
• Chen, Yuguo

Department of Study

Statistics

Discipline

Statistics

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Latent Class Model
• Bayesian
• magnetic resonance imaging (MRI)
• Co-segmentation
• Markov Random Field
• Market Segmentation
• Sparsity
• Collaborative
• Multiple Partition Process
• Dirichlet Process
• Indian Buffet Process
• Array Comparative genomic hybridization (ACGH)

Abstract

The latent class model (LCM) is a statistical method that introduces a set of latent categorical variables. The main advantage of LCM is that conditional on latent variables, the manifest variables are mutually independent of each other. In some scenarios, the LCM makes the modeling or computation feasible. In some other scenarios, the latent variables themselves are key. In the past a few decades, LCM has been widely applied to many areas such as Engineering, Medicine, Biology and Marketing. In this paper, several LCMs are developed in Bayesian framework to address new challenges in different applications. The first work is about the MR image segmentation. For MR images, we usually need to simultaneously segment multiple images, which are believed to have similar segmentation results. In our co-segmentation model, a Markov random field prior is utilized to encourage the information sharing. Clustering is usually regarded as an unsupervised problem. In our second work, we extend the clustering into supervised setting. This supervised clustering is evaluated in the application of market segmentation. In our third work, we relax the all-feature-in and all-object-in assumptions of the existing clustering approaches and propose a novel model called Multiple Partition Process (MPP) to obtain multiple clustering structures from the data. This MPP model is applied into the clustering of the breast cancer microarray data. In the last part of this paper, our future work is represented.

Graduation Semester

2011-05

Permalink

http://hdl.handle.net/2142/24418

Copyright and License Information

Copyright 2011 Jianfeng Xu

Owning Collections