Hierarchical Bayesian Neural Network for Correlated Gene Expression Temporal Pattern Classification

• Activity Number: 125
• Type: Contributed
• Date/Time: Monday, August 4, 2003 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Stat. Sciences
• Author(s): Yulan Liang*+ and Ebenezer George and Arpad Kelemen
• Companies: State University of New York at Buffalo and University of Memphis and University of Mississippi
• Keywords: hierarchical Bayesian neural networks ; correlated weights ; regularization ; gene expression temporal patterns ; hybrid Monte Carlo Markov chain

Abstract

An important issue in gene expression temporal patterns' analysis is the correlation structure of multidimensional microarray data. We propose a Hierarchical Bayesian Neural Network model with correlated weight structure and Bayesian regularization algorithm to characterize multiple gene functional temporal patterns. With the hierarchical Bayesian setting, the network parameters and hyperparameters were simultaneously optimized. Hybrid Markov chain Monte Carlo was employed for the network parameters learning in order to speed up the exploration of the parameter space. Gibbs sampling was applied to get hyperparameter estimation. Results show that the proposed model and algorithm can capture the gene expression temporal patterns despite the high noise levels, the highly correlated attributes, the overwhelming interactions, and other complex features that exist in the microarray data. The proposed model can deal with overtraining problems and improve the generalization performance. The performance of the model was also compared to the independent weights case and also to other popular machine learning methods such as Nearest Neighbor, Support Vector Machine, and Self-Organized Map.