Feature selection and classification by minimizing overlap degree for class-imbalanced data in metabolomics

Abstract

Learning from class-imbalanced data has gained increasing attention in recent years due to the massive growth of skewed data across many scientific fields such as metabolomics. Some researches show that it is not the imbalance itself which hinders the classification performance, but class overlapping do play an important role in the performance degradation when associated to class-imbalance. So alleviating the overlapping of the imbalanced data might be an effective way to improve the performance in class-imbalance learning. In this study, we propose two feature selection algorithms that aim to minimize the overlap degree between the majority and the minority, which is based on a simple assumption that decreasing overlap degree of a data set makes it more separable. The proposed MOSNS and MOSS methods are built via sparse regularization techniques. Simulation results indicate that our algorithms is effective in recognizing key features and control false discoveries for class-imbalance learning. Four class-imbalanced metabolomics data sets are also employed to test the performance of our algorithm, and a comparison with accuracy (ACC)-based and ROC-based selection procedures is performed. The result shows that our algorithms are highly competitive and can be an alternative feature selection strategy in class-imbalance learning.