关键词: 过采样, 类内不平衡, 少数类, 多样性, SMOTE算法, DBSCAN算法

Abstract:

For conventional oversampling algorithms, for example, SMOTE （Synthetic Minority Over-sampling Technique）, there are several problems such as ignoring within-class imbalances, extending the classification regions of minority class and synthesizing highly similar samples. Based on the comprehensive consideration of within-class imbalance and synthetic samples in diversity, an oversampling algorithm, which is a hybrid of DBSCAN and improved SMOTE （DB-MCSMOTE）, is proposed. It utilizes the DBSCAN algorithm to cluster the minority class samples. According to the proposed cluster density distribution function, the cluster density and sampling weight of each cluster are calculated. The MCSMOTE algorithm is adopted to oversample on the lines of the location-distant minority class samples in each cluster, the diversity of synthetic samples is improved and a new balanced dataset between and within classes is obtained. Experiments on a two-dimensional synthesis data set and nine UCI data sets show that DB-MCSMOTE can effectively improve the classification performance of the classifier for the minority class samples and the overall data set.

Key words: oversampling, within-class imbalance, minority class, diversity, Synthetic Minority Over-sampling Technique（SMOTE） algorithm, Density-Based Spatial Clustering of Applications with Noise（DBSCAN） algorithm