The class imbalance problem is one of the challenging issues in various machine learning applications. This problem occurs when the number of instances of a class is much smaller (or larger) than those of the other classes. To handle the imbalanced classification problems, many useful approaches have been developed, for example, synthetic minority oversampling technique (SMOTE). However, the SMOTE is often sensitive to the predetermined \(k\) value, i.e., the number of nearest neighbors used to generate the synthetic instances. For example, if the \(k\) value is moderately large, some of the synthetic instances generated by the SMOTE would be located near a decision boundary or even within the majority class area and thus these can be treated as unnecessary noisy instances. Thus, in this study, we propose an efficient hybrid resampling method based on instance density called IDHR (Instance Density-based Hybrid Resampling) to improve the classification performance by generating instances that are closer to the minority class than the majority class while avoiding generation of noisy instances. For this, we first apply the instance density-based oversampling (IDO) technique to generate new synthetic instances. And then, we eliminate some of the synthetic instances that are close to the decision boundary and determine the number of the synthetic instances among the retained synthetic ones which can be eliminated based on maximum of the distances from all the synthetic instances to the minority class instances and minimum of the distances from all the synthetic instances to the majority class instances as well as classification performances. To demonstrate the effectiveness of the proposed resampling method, comprehensive experiments are conducted on sixteen imbalanced datasets with considering three classifiers, i.e., C4.5 decision tree algorithm, support vector machine (SVM), and multi-layer perceptron neural network (MLP-NN). Through the experimental analysis, it is shown that the proposed resampling method outperforms the traditional oversampling methods with respect to AUC and F-measure for most of the imbalanced datasets regardless of classifiers.
