Feature Selection for High-Dimensional Genomic Microarray Data….pdfVIP

Feature Selection for High-Dimensional Genomic Microarray Data Eric P. Xing† epxing@cs.berkeley.edu Michael I. Jordan†‡ jordan@cs.berkeley.edu Richard M. Karp† karp@cs.berkeley.edu †Division of Computer Science, University of California, Berkeley, CA 94720 ‡Department of Statistics, University of California, Berkeley, CA 94720 Abstract ture selection and regularization methods. We report on the successful application of Feature selection methods have received much atten- feature selection methods to a classifica- tion in the classification literature (Kohavi John, tion problem in molecular biology involving 1997; Langley, 1994), where two kinds of meth- only 72 data points in a 7130 dimensional ods have generally been studied—filter methods and space. Our approach is a hybrid of filter wrapper methods. The essential difference between and wrapper approaches to feature selec- these approaches is that a wrapper method makes tion. We make use of a sequence of simple use of the algorithm that will be used to build the filters, culminating in Koller and Sahami’s final classifier, while a filter method does not. Thus, (1996) Markov Blanket filter, to decide on given a classifier C , and given a set of features F , particular feature subsets for each subset a wrapper method searches in the space of subsets cardinality. We compare between the re- of F , using cross validation to compare the perfor- sulting subset cardinalities using cross val-