An Improved Random Forest Classifier for Imbalanced Learning

Abstract

There are many application scenarios involving imbalanced datasets, whereas many traditional machine learning methods have limited ability to adapt to this kind of data. These methods usually have a bias to identify the majority classes while the minority classes are more important in many cases. In this study, we propose a variant of the completely random forest called HCRF. To improve the classification performance of imbalanced data, we introduced 2 mechanisms: random hybrid-resampling and a cost function that focuses on the minority classes. Verified on several imbalanced datasets, HCRF outperforms all comparison methods, demonstrating excellent performance on imbalanced learning.

Topics & Concepts