SMOTE-NaN-DE: Addressing the noisy and borderline examples problem in imbalanced classification by natural neighbors and differential evolution

摘要

Learning a classifier from class-imbalance data is an important challenge. Among existing solutions, SMOTE is one of the most successful methods and has an extensive range of practical applications. The performance of SMOTE and its extensions usually degrades owing to noisy and borderline examples. Filtering-based methods have been developed to address this problem but still have the following technical defects: (a) Error detection techniques heavily rely on parameter settings; (b) Examples detected by error detection techniques are directly eliminated, leading to deviation of obtained decision boundary and class imbalance again. To advance the state of the art, a novel filtering based oversampling method called SMOTE-NaN-DE is proposed in this paper. In SMOTE-NaN-DE, a SMOTE-based method is first used to generate synthetic samples and improve original class-imbalance data. Secondly, an error detection technique based on natural neighbors is used to detect noisy and borderline examples. Thirdly, the differential evolution (DE) is used to optimize and change iteratively the position (attributes) of found examples instead of eliminating them. The main advantages of SMOTE-NaN-DE are that (a) It can improve almost all of SMOTE-based methods in terms of the noise problem; (b) Error detection technique is parameter-free; (c) Examples found by error detection technique are optimized by the differential evolution rather than removed, which keeps imbalance ratio and improve the boundary; (d) It is more suitable for data sets with more noise (especially class noise). The effectiveness of the proposed SMOTE-NaN-DE is validated by intensive comparison experiments on artificial and real data sets. (C) 2021 Elsevier B.V. All rights reserved.