Erik Andries
Research Assistant
  Department of Mathematics and Statistics,
  Albuquerque High Performance Computing Center (AHPCC) &
  UNM Cancer Research Center
The University of New Mexico
Albuquerque, NM 87131
Phone: 505-277-6887, Fax: 505-277-8235
Email: andriese@math.unm.edu, andriese@ahpcc.unm.edu

   "Computationally Tractable Gene Selection
          Methods for Cancer Classification"

  Classification of patients into cancer subtypes or other medically
  relevant categories, from a numerical linear algebra perspective, is
  an ill-posed problem with respect to certain classes of classification
  algorithms (e.g., parametric hyperplane/hypersurface classifiers such
  as Fisher's Linear Discriminant (FLD) or Regularized Discriminant
  Analysis (RDA)).  Furthermore, the process of gene selection (i.e.
  identifying highly-discriminating subsets of genes that are the most
  responsible for the class distinctions of interest) can be
  computationally burdensome, requiring high performance computing--both
  serial and parallel.

  This paper/poster surveys
  a) a variety linear and nonlinear classification algorithms, e.g.
     -- large-margin classifiers such as support vector machines and
        boosting
     -- parametric hyperplane/hypersurface classification algorithms
        such as FLD and RDA,
     and
  b) strategies to make these algorithms effective and efficient with
     respect to both classification and gene selection, e.g.
     -- gradient-based methods,
     -- linear and nonlinear programming approaches,
     -- regularization and generalized inverses. 

  These methodologies will be benchmarked on both publically available
  microarray data sets and from leukemia microarray data sets from the
  UNM Cancer Research Center.        
====