An Improved SMOTE Algorithm for Enhancing Classification Performance in Imbalanced Adolescent Mental Health Datasets
Bin Pan, Haiqing Zhang, DaiWei Li, Xi Yu, Bo Cai, Junyu Deng · 2023
In order to address the performance bias in modeling on imbalanced datasets of adolescent mental health issues, as well as the issues introduced by existing oversampling methods that do not fully consider sample distribution and may introduce noise, a novel improved SMOTE algorithm called HD-SMOTE is proposed. This algorithm utilizes the concept of natural neighbors and the principle of information entropy to obtain the natural neighbor density of minority class samples and the density of sample boundaries, and then weights the two densities to obtain adaptive mixed density weights. Finally, these weights are used for adaptive synthetic oversampling. Experimental comparisons were conducted on six publicly available imbalanced datasets against four other classical algorithms. The results demonstrate that the proposed algorithm exhibits significant improvements in terms of AUC, F1, and G-mean. Furthermore, the combination of HD-SMOTE with the CatBoost algorithm shows strong performance in identifying adolescent mental health issues compared to two other ensemble learning methods, highlighting its practical value.