Reproduce “A systematic study of the class imbalance problem in convolutional neural networks” on OpenResearch

About this paper

Mateusz Buda, Atsuto Maki, Maciej A. Mazurowski

This systematic study by Mateusz Buda and colleagues investigates the class imbalance problem in Convolutional Neural Networks, empirically demonstrating its detrimental effect on performance. The research concludes that random minority oversampling is the most robust and effective mitigation strategy, consistently improving ROC AUC across various datasets without causing overfitting, and highlights that combining oversampling with thresholding can further enhance overall accuracy.

Key insights

• Class imbalance has a substantial detrimental effect on CNN classification performance, with the negative impact being stronger for more complex tasks.
• Random minority oversampling is the most consistently effective method for mitigating class imbalance in CNNs with respect to ROC AUC.
• Oversampling does not cause overfitting in Convolutional Neural Networks, a finding that contradicts observations from some classical machine learning models.
• Thresholding, when combined with oversampling, can effectively improve overall accuracy, serving as a valuable approach when this metric is prioritized.

Summary by alphaXiv.