Multi-class feature selection using Pairwise-class and All-class techniques

Feature selection has been a key technique in massive data processing, e.g. microarray data analysis with few samples but high dimensions. One common problem in multi-class data analysis is the unbalanced recognition accuracies among classes, which leads to poor system performance. One main reason is that most feature selection methods focus on the performance of whole dataset while pay little attention to single class (especially the minority class). In this paper, a novel hybrid feature selection method with Pairwise-class and All-class techniques (namely FSPA) is proposed to remedy the problem. Strategy of round-robin is embedded into FSPA to reduce the bias among classes. Experimental results on four public microarray datasets show that FSPA helps to achieve higher classification accuracy and balance the performance among classes.