ND-S: an oversampling algorithm based on natural neighbor and density peaks clustering

There are a large number of imbalanced classification problems in the real world. Due to the imbalance in the amount of data and the complex nature of the distribution, the minority class samples are difficult to be classified correctly. Oversampling techniques balance the data set by generating minority class samples; however, current clustering-based oversampling techniques are limited by hyperparameters, sample selection, and other issues that affect the final classification performance. In this paper, we propose an oversampling algorithm based on natural neighbor and density peaks clustering (ND-S). ND-S is divided into three steps. Firstly, the natural neighbor algorithm is used to find and filter noises and outliers. Secondly, the density peaks clustering is improved by natural neighbor-based nonparametric adaptive, which clusters all samples and leaves the clusters that meet the conditions. Finally, sampling weights are assigned to each cluster, and the minority class of samples suitable for oversampling is selected for synthetic minority oversampling (SMOTE) by calculating the local sparsity of the samples. Experiments on 18 imbalanced data sets show that ND-S is effective for the imbalanced classification problem, and its classification performance is generally better than other 8 comparison algorithms.