Ensemble-based Noise Filter for Noisy and Imbalanced Data

Class Imbalance problems have an unequal distribution of classes within a dataset. The noise filter is a preprocessing technique that aims to identify and delete noisy examples in the training data to improve the performance of an induced classifier. In our previous work, we proposed a novel method deletes noisy examples, located inside the borderline region, from the majority class. The method combines the advantages of thresholding with noise filtering. In that method, the identification of noise is based on classification filter approach. One of the main drawbacks of this approach is that classifiers are organized in parallel way so they cannot benefit each other through noise identification. In this paper, we evaluate our proposed method by investigating the use of ensemble filter approach to identify noisy examples. The extended method has been assessed on 14 real imbalanced datasets. The results show that our proposal is significant in identifying noise and performs better in term of AUC scores based on random forest.