TDMO: Dynamic multi-dimensional oversampling for exploring data distribution based on extreme gradient boosting learning

The synthetic minority oversampling technique (SMOTE) is the most general and popular solution for imbalanced data. Although SMOTE is effective in solving the class imbalance problem in most cases, it insufficiently exploits the data prior distribution. Additionally, most existing SMOTE variants randomly produce new instances between a minority sample and its nearest neighbors, which carries the risk of noise propagation. To address this, in this paper, local distribution trust estimation based on extreme gradient boosting (XGBoost) and dynamic multi-dimensional oversampling (TDMO) is proposed as a novel approach to exploring data distributions. First, undersampling and XGBoost techniques are introduced to train multiple balanced subsets to identify the internal structure of the original data and obtain the classification prediction accuracy of each instance, called the confidence level (CL). Then, instances with low CL (i.e., noise) are filtered out, and the densities of the two classes in the neighborhood of the non-noise instances are evaluated to create candidate samples to expand the diversity of the minority class. Finally, the minority class is enhanced by combining multiple samples in a multi-dimensional feature space. Extensive experimental results demonstrate that TDMO outperformed the comparative oversampling methods clearly and obtained the optimal classification results.

Automated summary

This paper proposes a novel approach based on XGBoost and TDMO to address the issue of data distribution. By training multiple balanced subsets, filtering noise, and combining multiple samples, the diversity of the minority class is expanded, resulting in superior classification results compared to other methods.