OC-SMOTE-NN: A Deep Learning-based Approach for Imbalanced Classification

Navodika Karunasingha, Buddhi G. Jayasekara, Asela Hevapathige

Abstract: We propose a deep learning-based over-sampling approach to overcome the class imbalance problem. Our model consists of three major components: 1) A variant of SMOTE algorithm with learnable parameters; 2) A deep-learning model as a discriminator; 3) A loss function with an orthogonality-based regularization term. Contrary to traditional over-sampling approaches with manual fine-tuning, our model is capable of performing adaptive over-sampling. The experimental results demonstrate that our model considerably outperforms the baselines with up to a 9% increase in F1-score or provides comparable performance.