NI-MWMOTE: An improving noise-immunity majority weighted minority oversampling technique for imbalanced classification problems

Oversampling techniques have been favored by researchers because of their simplicity and versatility in dealing with imbalanced classification problems. For oversampling techniques appeared in recent years (e.g. Majority Weighted Minority Oversampling Technique (MWMOTE)), noise processing plays an important role. This is because the processing of noise directly affects the distribution of new synthetic instances. MWMOTE and many other oversampling techniques use knn based noise processing method. While the knn method can effectively handle partial noise when the neighborhood parameter k value is reasonable, it may lead to under-recognition or over-recognition without prior experience. Therefore, we propose an improving noise-immunity majority weighted minority oversampling technique abbreviated NI-MWMOTE. NI-MWMOTE uses an adaptive noise processing scheme, which combines Euclidean distance and neighbor density to rank the probability that suspected noise (knn method) is true noise, and then adaptively selects the best noise processing scheme through iteration and misclassification error. Then, aggregative hierarchical clustering (AHC) method is used to cluster minority instances. And, in each sub-cluster, the sampling size of new samples is adaptively determined by classification complexity and cross-validation. NI-MWMOTE not only avoids the generation of new noise, but also effectively overcomes both between-class imbalances and within-class imbalances. Results demonstrate that NI-MWMOTE achieves significantly better results in most imbalanced datasets than eight popular oversampling algorithms. (c) 2020 Elsevier Ltd. All rights reserved.