A hybrid ensemble method based on double disturbance for classifying microarray data.

ABSTRACT

Microarray data has small samples and high dimension, and it contains a significant amount of irrelevant and redundant genes. This paper proposes a hybrid ensemble method based on double disturbance to improve classification performance. Firstly, original genes are ranked through reliefF algorithm and part of the genes are selected from the original genes set, and then a new training set is generated from the original training set according to the previously selected genes. Secondly, D bootstrap training subsets are produced from the previously generated training set by bootstrap technology. Thirdly, an attribute reduction method based on neighborhood mutual information with a different radius is used to reduce genes on each bootstrap training subset to produce new training subsets. Each new training subset is applied to train a base classifier. Finally, a part of the base classifiers are selected based on the teaching-learning-based optimization to build an ensemble by weighted voting. Experimental results on six benchmark cancer microarray datasets showed proposed method decreased ensemble size and obtained higher classification performance compared with Bagging, AdaBoost, and Random Forest.