Synthetic Information Toward Maximum Posterior Ratio for Deep Learning on Imbalanced Data
Hung T. Nguyen, J. Morris Chang · IEEE Transactions on Artificial Intelligence · 2023
This work explores how class-imbalanced data affects deep learning and proposes a data balancing technique for mitigation by generating more synthetic data for the minority class. In contrast to random-based oversampling techniques, our approach prioritizes balancing the most informative region by finding high entropy samples. This approach is opportunistic and challenging because well-placed synthetic data points can boost machine learning algorithms’ accuracy and efficiency, whereas poorly-placed ones can cause a higher misclassification rate. In this study, we present an algorithm for maximizing the probability of generating a synthetic sample in the correct region of its class by placing it toward maximizing the class posterior ratio. In addition, to preserve data topology, synthetic data are closely generated within each minority sample neighborhood. Overall, experimental results on forty-one datasets show that our technique significantly outperforms experimental methods in terms of boosting deep-learning performance.