Deep learning models often exhibit bias when trained on imbalanced datasets, tending to favor majority classes. While Deep Synthetic Minority Over-sampling Technique (DeepSMOTE) mitigates this challenge by generating synthetic minority class images within an autoencoder’s latent space, the proposed architecture achieves further performance enhancement through the optimization of the latent space structure. This strategy introduces a composite loss function that minimizes intra-class distances while maximizing inter-class distances by incorporating class centroid information into the autoencoder training process. The approach was evaluated on six benchmark image datasets (MNIST, Fashion-MNIST, EMNIST, SVHN, GTSRB, and CIFAR10) with severe class imbalance ratios of 750:1, using five-fold cross-validation. Results demonstrate that Deep Class-Latent Synthetic Minority Oversampling Technique (DeepCLSMOTE) consistently outperforms baseline methods, including Balancing Generative Adversarial Network (BAGAN) and DeepSMOTE across all evaluation metrics, achieving statistically significant improvements in macro-average accuracy, precision, recall, and F1-measure. The enhanced performance is attributed to improved discriminative feature extraction in the optimized latent space, resulting in superior classification performance on highly imbalanced image datasets, particularly for critical minority classes.
Homjandee et al. (Fri,) studied this question.