Exploring an iterative feature selection technique for highly imbalanced data sets

The quality of a classification model is affected by two factors in a training data set: (1) the presence of excessive features and (2) the presence of imbalanced distributions between two classes in a binary classification problem. This paper presents an iterative feature selection method to deal with these two problems. The proposed method consists of an iterative process of data sampling followed by feature ranking and finally aggregating the results generated during the iterative process. In this study, we investigate a number of feature ranking techniques and a data sampling method with two different post-sampling proportions between the two classes. We compare the iterative feature selection technique to the one where a data sampling and a feature ranking technique are used together but only once (without iteration). The empirical study is carried out on two groups of highly imbalanced data sets from a real-world software system. The results demonstrate that our proposed iterative feature selection technique performs on average better than the method without iteration.