An oversampling method for imbalanced dataset based on sparsity and boundary degree

In order to improve the classification accuracy of minority samples in imbalanced dataset, we proposed a novel oversampling method, HRSB-SMOTE (HDBSCAN-Ratio-Sparsity-Boundary-SMOTE), which is based on HDBSCAN clustering and SMOTE, and combined with cluster ratio, sparsity and boundary degree of minority samples. Firstly, we implemented HDBSCAN clustering on the minority samples and removed the noisy samples by grade of membership. Then, according to the cluster ratio and sparsity of the minority samples in each cluster, we ascertained the number of needed synthetic samples for such cluster. Afterwards, we determined the number of needed synthetic samples for each minority sample based on the boundary degree of it in each cluster. Finally, we synthesized minority sample by using SMOTE according to the aforementioned number. Our proposed method not only can preserve the distribution characteristics of original data, but also can reinforce the decision boundary and avoid the generation of noisy samples. The experimental results on 13 real-world datasets from UCI Repository show that the performance of HRSB-SMOTE method is superior to other six popular oversampling methods (such as SMOTE, Borderline-SMOTE, ANASYN, etc.) in terms of F-measure, G-mean, and Acc on most datasets. Compared with SMOTE, Borderline-SMOTE, and k-means-SMOTE, the F-measure value of HRSB-SMOTE method on winequality-red-8vs6 dataset with higher IR (Imbalance Ratio) and with KNN classifier is improved by 10.79%, 1.16%, and 7.89%, respectively. HRSB-SMOTE method effectively handles the imbalanced problem between and within classes.