Feature Selection using One Class SVM: A New Perspective

Feature selection is an important task for mining useful information from datasets in high dimensions, a typical characteristic of biology domains such as microarray datasets. In this paper, we present an altogether new perspective on feature selection. We pose feature selection as a one class SVM problem of modeling the space in which features can be represented. We show that finding the support vectors in our one class formulation is tantamount to performing feature selection. Further, we show that our formulation reduces to the standard QPFS formulation in the dual problem space. Not only our formulation gives new insights into the task of feature selection, solving it directly in the primal space can give significant computational gains when the number of the samples is much smaller than the number of features. We validate our thesis by experimenting on three different microarray datasets.