Improving the performance of SVM-RFE on classification of pancreatic cancer data

Feature selection is a key step in classification of high-dimensional data, especially gene expression microarray data with many thousands of features. As a wrapper method, Support Vector Machine-Recursive Feature Elimination (SVM-RFE) is one of the most powerful feature selection techniques. Although SVM-RFE can remove irrelevant features effectively, it cannot deal with most of the redundant features. To overcome this drawback, this paper develops a new feature selection method, the core of which is removing redundant features based on the correlation among features before using SVM-RFE. We test the proposed method on the pancreatic cancer microarray dataset. The experimental results show that our method performs much better than the baseline SVM-RFE technique in terms of classification accuracy. To improve the class-wise classification accuracies, radial basis function (RBF) kernel is also incorporated.