Ensemble learning-based filter-centric hybrid feature selection framework for high-dimensional imbalanced data

In recent years, research on feature selection for high-dimensional imbalanced data has attracted a considerable amount of attention. The filter-wrapper hybrid method, which is a conventional method of feature selection for high-dimensional data, aims to reduce excessive computational time. On the other hand, ensemble learning-based feature selection, even though it has a high level of computational complexity, focuses exclusively on the discovery of robust features. From this perspective, combining these two feature selection methods is not easy. However, a combined method is essential to advancing machine learning research that addresses real-world problems. We propose an filter-centric hybrid method based on ensemble-learning that can select the best feature subset for high-dimensional imbalanced data. The basic concept of the proposed method is to design a feature evaluation scheme based on the filter method and to apply ensemble learning with reasonable computational time. To achieve this objective, our innovative method utilizes predictions produced by multiple classifiers as inputs of the feature evaluation function. As a result, it can reflect the predictive performance of the classifiers and overcome the low performance of selected features by filter methods. In addition, it can find robust features simultaneously. To demonstrate the superiority of the proposed method, we perform various experiments using 14 experimental datasets that consist of low-dimensional balanced, high-dimensional balanced, and high-dimensional imbalanced datasets. Finally, we compare the proposed method with state-of-the-art feature selection methods.