Symmetric uncertainty class-feature association map for feature selection in microarray dataset

Abstract

For a huge number of features versus a small size of samples, feature selection methods are useful preprocessing approaches that could eliminate the irrelevant and redundant features from the final feature subset. One of the recent research areas in feature selection is DNA microarray that the number of dimensions increase fast and requires further research in the field of feature selection. Modeling the feature search space as a graph leads to improving the visualizing of features and using graph theoretic concepts in the feature selection process. In this paper, a filer-based feature selection algorithm using graph technique is proposed for reducing the dimension of dataset named as Symmetric Uncertainty Class-Feature Association Map feature selection (SU-CFAM). In the first step, it uses the Symmetric Uncertainty concept for visualizing the feature search space as a graph. After clustering the graph into several clusters using a community detection algorithm, SU-CFAM constructs an adjacency matrix for each cluster and the final subset is selected by using the concept of maximal independent set. The performance of SU-CFAM has been compared with five well-known feature selection approaches using three classifiers including SVM, DT, NB. Experiments on fifteen public DNA microarray datasets show that SU-CFAM can achieve a better classification performance compared with other methods.