Evaluation of Gene Selection Using Support Vector Machine Recursive Feature Elimination

2 One advantage of the microarray technique is that it allows scientists to explore the ex-pression of thousands of genes in a single expression microarray. This creates a set of high dimensional data with more than fifty thousand features. The price of a microarray is still high, thus we may not have many samples in a single microarray experiment. So this small data set with high dimensional feature space requires a special technique for its analysis. One common requirement is a small subset of genes so that we can build the most accurate classifier for predicting in-coming data. Working in high dimensional space requires excessive computational cost. The main problem of small data set is overfitting. Feature ranking reduces the computational cost and overfitting. Guyon et al.[7] proved that support vector recursive feature elimination performs better in gene selection compared to other existing methods. Recursive feature elimination is a top-down wrapper method using weight vector as feature ranking criterion. In this experiment, we explore the interaction of the support vector machine correctness with the number of genes and the rate of feature elimination. The result shows that there is no significance in the rate of feature elimination when the surviving subset is large, thus we can eliminate the genes quickly in feature selection. The smaller data sets still need more investigation.