Stable variable selection of class-imbalanced data with precision-recall criterion

• Published in: Chemometrics and intelligent laboratory systems 2017-12, Vol.171, p.241-250
• Main Authors: Fu, Guang-Hui, Xu, Feng, Zhang, Bing-Yang, Yi, Lun-Zhao
• Format: Article
• Language: English
• Subjects: Class-imbalanced data; Precision-recall curve; Sparse regularization logistic regression; Stable variable selection; Subsampling
• ISSN: 0169-7439; 1873-3239
• DOI: 10.1016/j.chemolab.2017.10.015

Screening important variables for class-imbalanced data is still a challenging task. In this study, we propose an algorithm for stably selecting key variables on class-imbalanced data based on the precision-recall curve (PRC), where the PRC is utilized as the assessment criterion in the model building stage, and sparse regularized logistic regression combined with subsampling (SRLRS) is designed to perform stable variable selection. Considering the characteristic of class-imbalanced data, we also proposed classification-based partition for cross validation, as well as leaving half of majority observations out and leaving one minority observation out (LHO-LOO) for subsampling. Simulation results and real data showed that our algorithm is highly suitable for handling class-imbalanced data, and that the PRC can be an alternative evaluation criterion for model selection when handling class-imbalanced data. •Precision-recall curve (PRC) as a criterion for variable selection of class-imbalanced data.•A novel algorithm (SRLRS) is proposed for dealing with class-imbalanced data.•A novel subsampling (LHO-LOO) strategy for class-imbalanced data is designed for stable variable selection.•Sparse regularized methods are successfully used for class-imbalanced data.