Open Access

ARTICLE

Cost-Sensitive Dual-Stream Residual Networks for Imbalanced Classification

by Congcong Ma^1,2, Jiaqi Mi¹, Wanlin Gao^1,2, Sha Tao^1,2,*

1 College of Information and Electrical Engineering, China Agricultural University, Beijing, 100083, China
2 Key Laboratory of Agricultural Informatization Standardization, Ministry of Agriculture and Rural Affairs, China Agricultural University, Beijing, 100083, China

* Corresponding Author: Sha Tao. Email:

Computers, Materials & Continua 2024, 80(3), 4243-4261. https://doi.org/10.32604/cmc.2024.054506

Received 30 May 2024; Accepted 07 August 2024; Issue published 12 September 2024

Abstract

Imbalanced data classification is the task of classifying datasets where there is a significant disparity in the number of samples between different classes. This task is prevalent in practical scenarios such as industrial fault diagnosis, network intrusion detection, cancer detection, etc. In imbalanced classification tasks, the focus is typically on achieving high recognition accuracy for the minority class. However, due to the challenges presented by imbalanced multi-class datasets, such as the scarcity of samples in minority classes and complex inter-class relationships with overlapping boundaries, existing methods often do not perform well in multi-class imbalanced data classification tasks, particularly in terms of recognizing minority classes with high accuracy. Therefore, this paper proposes a multi-class imbalanced data classification method called CSDSResNet, which is based on a cost-sensitive dual-stream residual network. Firstly, to address the issue of limited samples in the minority class within imbalanced datasets, a dual-stream residual network backbone structure is designed to enhance the model’s feature extraction capability. Next, considering the complexities arising from imbalanced inter-class sample quantities and imbalanced inter-class overlapping boundaries in multi-class imbalanced datasets, a unique cost-sensitive loss function is devised. This loss function places more emphasis on the minority class and the challenging classes with high inter-class similarity, thereby improving the model’s classification ability. Finally, the effectiveness and generalization of the proposed method, CSDSResNet, are evaluated on two datasets: ‘DryBeans’ and ‘Electric Motor Defects’. The experimental results demonstrate that CSDSResNet achieves the best performance on imbalanced datasets, with macro_F1-score values improving by 2.9% and 1.9% on the two datasets compared to current state-of-the-art classification methods, respectively. Furthermore, it achieves the highest precision in single-class recognition tasks for the minority class.

---

Keywords

---