Siamese networks with an online reweighted example for imbalanced data learning. (December 2022)

• Record Type: Journal Article
• Authors: Zhao, Linchang; Shang, Zhaowei; Tan, Jin; Zhou, Mingliang; Zhang, Mu; Gu, Dagang; Zhang, Taiping; Tang, Yuan Yan
• Is Part Of: Pattern recognition. Volume 132(2022)
• Publication Date: 2022-12
• Subjects: Few-shot learning -- Reweighted example learning -- Data mining -- Imbalanced learning; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2022.108947
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A unified approach that combines few-shot learning and reweighted example learning called ORESCNN is used for classifying unbalanced datasets. An ORE algorithm is introduced to handle the problem of data with a class-imbalanced distribution, which alleviates the influence of imbalanced data on model performance by reweighted example learning. Extensive experiments are conducted on experimental databases, and the results show that the proposed approach favourably performs against baseline methods. Abstract: One key challenging problem in data mining and decision-making is to establish a decision support system based on unbalanced datasets. In this study, we propose a novel algorithm to handle unbalanced learning problems that integrates the advantages of Siamese convolutional neural networks (SCNN) and the online reweighted example (ORE) algorithm into a unified method. First, the SCNN model is established for learning and extracting deep feature features at different levels. Second, the ORE algorithm is used to address the problem of data with a class-imbalanced distribution. Compared with baseline approaches, the experimental results show that our proposed method substantially enhances the performance of both within-project defect prediction and cross-project defect prediction.