Distance-based arranging oversampling technique for imbalanced data

Class imbalance data sets are common in a vast variety of real-world application areas. Synthetic minority oversampling technique (SMOTE) is an important technique for processing imbalanced data sets. SMOTE requires the user to preset the number of nearest neighbor instances before synthesizing instances, which is often difficult to choose accurately. Moreover, SMOTE is easy to synthesize minority instances in the majority areas, which leads to the performance degradation of the classifier. To address these issues, in this paper, a novel distance-based arranging oversampling (DAO) technique is proposed. DAO can effectively prevent users from selecting inaccurate hyperparameters, and DAO can be used as an alternative algorithm to replace the SMOTE-based oversampling technique. We further filter the synthesized instances by setting appropriate conditions to avoid generating minority instances in the majority domain. In our experiments, we collect 25 public benchmark data sets from the KEEL database and HDDT database, and apply CART and ID3 classification models on the oversampling training set of each data set to assess our DAO technique. Under the two evaluation metrics, F-measure and kappa, compared with the state-of-the-art oversampling techniques, our proposed method is superior or partially superior to them.