LASSO-Patternsearch for Multivariate Bernoulli Observations with Applications

• Activity Number: 41
• Type: Contributed
• Date/Time: Sunday, August 1, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Bin Dai*+ and Stephen Wright and Xiwen Ma and Grace Wahba
• Companies: University of Wisconsin-Madison and University of Wisconsin-Madison and University of Wisconsin-Madison and University of Wisconsin-Madison
• Keywords: LASSO ; Multivariate Bernoulli ; Generalized linear model ; GACV

Abstract

The LASSO-Patternsearch algorithm is proposed to identify patterns efficiently that arise from the log linear expansion of the multivariate Bernoulli distribution. The method is designed for cases in which there is a very large number of candidate patterns but it is believed that relatively few are important. LASSO is used to reduce the number of candidate patterns greatly, using a novel computational algorithm that can handle a large number of unknowns simultaneously. The joint distribution conditioned on the predictor variables is estimated and the log odds ratio is used to measure the association among outcome variables. A data-adaptive tuning procedure based on GACV, modified to act as a model selector is proposed. Simulation studies and application to benchmark data set are conducted to check the performance of the proposed method.