Imbalanced Multi-class Classification Based on Bi-Criterion Evolution Sampling and Ensemble Learning
Yun Yu, Lin Meng, Zixuan Wang, Yongqing Zhang · 2023
Imbalanced dataset is a common phenomenon in real word applications. However, most of current classification approaches usually assume that the dataset is balanced, which results in undesirable classification performance, especially in multi-class classification. Here, we propose an approach based on bi-criterion evolution sampling and ensemble learning (BCEC) for imbalanced multi-class classification. Specifically, BCEC employs an evolution algorithm based on both Pareto criterion and non-Pareto criterion to undersample from each majority class for building balanced datasets. Then, BCEC separately characterizes each balanced dataset by a multi-layer perceptron and utilizes the ensemble learning based on Bagging strategy to integrate the classification results. Extensive experiments over UCI datasets demonstrate that our approach significantly outperforms several state-of-the-art imbalanced classification methods in terms of G-means.