A fuzzy approach to clustering and selecting features for classification of gene expression data

Classification assigns a discrete value named label to each sample in a dataset with respect to its feature values. In this research, we aim to consider some datasets which contain a few samples whereas a huge amount of features are provided for each sample. Most of biological datasets such as micro-arrays has this property. A fundamental contribution of this article is a major extension of pervious works for crisp data clustering. The new approach is based on fuzzy feature clustering which is utilized to select the best features (genes). The proposed method has two advantages over the crisp method. Firstly, it leads to more stability and faster convergence; secondly, it improves the accuracy of the classifier using the selected features. Moreover, in this paper a novel method has been proposed for the discretization of continuous data using the Fisher criterion. In addition, a new method for initialization of cluster centers is suggested. The proposed method has achieved a considerable improvement compared with the crisp version. The leukemia dataset has been used to illustrate the effectiveness of the method.