Feature selection based on pairwise evalution

Feature selection is an important tool used in data reduction; it aims at improving efficiency in many machine-learning algorithms by choosing a small set of informative features among the whole dataset. Feature selection algorithms can be classified in three major categories: Filter, Wrapper and Embedded. In this paper, we proposed a new hybrid filter-wrapper algorithm of feature selection based on pairwise feature selection, which benefits from the speed up and the ease of use of filters and the good performance of wrappers. The presented algorithm used the decision tree to evaluate individual as well as pair features. The score is obtained using the Area under the ROC curve (AUC) of decision tree algorithm. Our method is compared with an algorithm of feature selection, which is based on pairwise evaluation as well. Six benchmark datasets are used in this work, which are available on UCI machine learning repository. The results demonstrate that the proposed method is capable of producing a good performance.