Comparison and Evaluation of the Combinations of Feature Selection and Classifier on Microarray Data

As gene chip technology is widely used in cancer research, a large number of valuable microarray data has been rapidly accumulated. These data have the characteristics of “high-dimensional small samples”, in which most genes are unrelated or redundant. For high-dimensional, small-sample, high-noise, and few-sample binary classification datasets, we explore which combination of feature selection method and classifier can achieve the relatively best prediction accuracy, while the number of features included is relatively low. We adopt the standard data analysis procedures: preprocessing the data set, using different feature selection methods to generate feature subsets, and applying different classifiers to predict each feature subset. The results are compared to find out which combination with the relatively high prediction accuracy and the relatively small number of features.