SIA-SMOTE: A SMOTE-based Oversampling Method with Better Interpolation on High-Dimensional Data by Using a Siamese Network

SMOTE is an effective method for balancing imbalanced datasets by interpolating between existing samples in the minority class. However, if the synthetic samples generated through interpolation are based on noisy data points, then they may also be noisy and can lead to overfitting and reduced performance on unseen data. This paper proposes a new method SIA-SMOTE, which uses SMOTE for oversampling the minority class and a siamese network for synthetic image selection. SIA-SMOTE also explores the decision boundary to better capture data distribution of the minority class. The proposed method has been compared to random oversampling, SMOTE, and ASN-SMOTE on MNIST, FMNIST, and three medical image datasets. The results show that SIASMOTE achieved the best overall performance in terms of three evaluation metrics.