{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/a-systematic-study-of-the-class-imbalance","title":"A systematic study of the class imbalance problem in convolutional neural networks","arxiv_id":"1710.05381","date":"2017-10-15","proceeding":null,"authors":["Mateusz Buda","Atsuto Maki","Maciej A. Mazurowski"],"abstract":"In this study, we systematically investigate the impact of class imbalance on\nclassification performance of convolutional neural networks (CNNs) and compare\nfrequently used methods to address the issue. Class imbalance is a common\nproblem that has been comprehensively studied in classical machine learning,\nyet very limited systematic research is available in the context of deep\nlearning. In our study, we use three benchmark datasets of increasing\ncomplexity, MNIST, CIFAR-10 and ImageNet, to investigate the effects of\nimbalance on classification and perform an extensive comparison of several\nmethods to address the issue: oversampling, undersampling, two-phase training,\nand thresholding that compensates for prior class probabilities. Our main\nevaluation metric is area under the receiver operating characteristic curve\n(ROC AUC) adjusted to multi-class tasks since overall accuracy metric is\nassociated with notable difficulties in the context of imbalanced data. Based\non results from our experiments we conclude that (i) the effect of class\nimbalance on classification performance is detrimental; (ii) the method of\naddressing class imbalance that emerged as dominant in almost all analyzed\nscenarios was oversampling; (iii) oversampling should be applied to the level\nthat completely eliminates the imbalance, whereas the optimal undersampling\nratio depends on the extent of imbalance; (iv) as opposed to some classical\nmachine learning models, oversampling does not cause overfitting of CNNs; (v)\nthresholding should be applied to compensate for prior class probabilities when\noverall number of properly classified cases is of interest.","url_abs":"http://arxiv.org/abs/1710.05381v2","url_pdf":"http://arxiv.org/pdf/1710.05381v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"a-systematic-study-of-the-class-imbalance","repo_url":"https://github.com/cogsci2/Pneumonia-XRay-Differentiation-from-Kaggle-Dataset","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}},{"paper_slug":"a-systematic-study-of-the-class-imbalance","repo_url":"https://github.com/ferhatkkochan/deeplearning-notes","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}},{"paper_slug":"a-systematic-study-of-the-class-imbalance","repo_url":"https://github.com/fkochan/deeplearning-notes","is_official":0,"mentioned_in_paper":0,"mentioned_in_github":1,"framework":"none","reach":{"status":"ok"}}],"tasks":[{"task_slug":"machine-learning","task_name":"BIG-bench Machine Learning"},{"task_slug":"classification","task_name":"General Classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"syntology_url":"https://syntology.ai/paper/1710.05381","atlas_url":"https://app.syntology.ai/?focus=1710.05381","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}