Journal article
Fuzzy data augmentation for handling overlapped and imbalanced data
Neural Information Processing, Vol.1516, pp.625-633
2021
Abstract
Class imbalance is a serious issue in classification as a traditional classifier is generally biased towards the majority class. The accuracy of the classifier could be further impacted in cases where additionally to the class imbalance, there are overlapped data instances. Further, data sparsity has shown to be a possible issue that may lead to non- invariance and poor generalisation. Data augmentation is a technique that can handle the generalisation issue and improve the regularisation of the Deep Neural Network (DNN). A method to handle both class overlap and class imbalance while also incorporating regularisation is proposed in this paper. In our work, the imbalanced dataset is balanced using SMOTETomek, and then the non-categorical attributes are fuzzified. The purpose of fuzzifying the attributes is to handle the overlapping in the data and provide some form of data augmentation that can be used as a regularisation technique. Therefore, in this paper, the invariance is achieved as the augmented data are generated based on the fuzzy concept. The balanced augmented dataset is then trained using a DNN classifier. The datasets used in the experiments were selected from UCI and KEEL data repositories. The experiments show that the proposed Fuzzy data augmentation for handling overlapped and imbalanced data can address the overlapped and imbalanced data issues, and provide regularisation using data augmentation for numerical data to improve the performance of a DNN classifier.
Details
- Title
- Fuzzy data augmentation for handling overlapped and imbalanced data
- Authors/Creators
- R. Dabare (Author/Creator) - Murdoch UniversityK.W. Wong (Author/Creator) - Murdoch UniversityM.F. Shiratuddin (Author/Creator) - Murdoch UniversityP. Koutsakis (Author/Creator) - Murdoch University
- Publication Details
- Neural Information Processing, Vol.1516, pp.625-633
- Publisher
- Springer Verlag
- Identifiers
- 991005542872907891
- Copyright
- © 2021 Springer Nature Switzerland AG
- Murdoch Affiliation
- School of Information Technology
- Language
- English
- Resource Type
- Journal article
- Additional Information
- Paper given @ International Conference on Neural Information Processing (ICONIP 2021): Neural Information Processing
Metrics
119 Record Views