Order-Preserving Factor Discovery with Misaligned Data

• Activity Number: 401
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 2:00 PM to 3:50 PM
• Sponsor: IMS
• Author(s): Arnau Tibau Puig+ and Ami Wiesel and Alfred O. Hero*
• Companies: University of Michigan and University of Michigan and University of Michigan
• Keywords: Factor analysis ; convex optimization ; sparse dictionary learning

Abstract

We present a method to discover factors that influence the temporal profiles of a large number of variables in a population of subjects undergoing a common treatment but whose responses may not be temporally aligned. Our main hypothesis is that there exists a subset of variables with similar but delayed profiles that obey some unknown but consistent precedence ordering relationship. The proposed method uses a linear factor model with additional degrees of freedom that account for each subject's inherent delays under an order-consistent constraint. We present an efficient algorithm to fit the model that is similar to sparse dictionary learning but incorporates a delay estimation step that aligns the factors to each subject's response. The method will be illustrated on a large gene expression dataset.