Multi-fidelity model based on synthetic minority over-sampling technique

Oversampling is a commonly employed technique to address class imbalance problems by equalizing the sizes of different data classes through the addition of the minority class data. The accuracy of the classification in the presence of class imbalance heavily relies on the quality of the generated minority class data. To ensure the quality of the input data for the classification model and enhance the stability of the oversampling approach, this study proposes a multi-fidelity model known as MFSMOTE, which is based on the synthetic minority group oversampling technique. MFSMOTE partitions the dataset into low-fidelity and high-fidelity data, enabling the interaction between them to yield a more robust classification model. Specifically, the low-fidelity data provide prior information to guide the high-fidelity data, which, in turn, is used to train the classification model. Finally, MFSMOTE is compared with 12 existing oversampling algorithms across 21 imbalanced datasets. The experimental results demonstrate that MFSMOTE serves as a promising novel oversampling method.