Feature Selection with SVM Reformulation and Expression-array-based Cancer Classification

• Activity Number: 384
• Type: Contributed
• Date/Time: Wednesday, August 11, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Sunhee K. Ro and Wei-min Liu*+
• Companies: Roche Molecular Systems, Inc. and Roche Molecular Systems, Inc.
• Keywords: microarray ; cancer classification ; feature selection ; support vector machine

Abstract

Gene expression microarray analysis is being used extensively for cancer classification and support vector machine (SVM) is one of the successful classification methods. While SVM can be applied to extremely high-dimensional data such as microarray, the removal of irrelevant and redundant features may produce better results. Classical wrapper methods such as the SVM RFE can be used for feature selection while avoiding the combinatorial explosion with greedy technique; however, it is slow because the classification performance should be calculated at every iteration to find the optimum number of features. Weston et al. (2003) proposed to reformulate the SVM problem as minimizing the number of nonzero elements in the weight vector in SVM decision function. The algorithm amounts to a modification of SVM with multiplicative rescaling of the training data at every iteration so that the weights of useless features approach 0 rapidly. We have extended this algorithm to classify the multiple subclasses of leukemia, which consists of the all-pairwise classifiers each of which was trained with the union of the genes selected pairwisely with the multiplicative rescaling.