A Modified Borderline Smote with Noise Reduction in Imbalanced Datasets

In the real world, noisy data brings tremendous challenges to data mining. Traditional classification methods are proven to be inadequate to assess the efficacy of the data mining methods while using noisy and imbalanced data. Therefore, preprocessing the imbalanced data is necessary before classification. But it's difficult to arrive at an appropriate classifier for minority class in the imbalanced data. This paper proposes the hybridization of two techniques, Noise reduction and oversampling techniques which only oversamples or strengthens the borderline minority class. The proposed technique is applied on 49 datasets at several imbalanced ratios. The Decision Tree, Gaussian Naive Bayes, Logistic Regression, Neural Network, Non-linear SVM, Random Forest, and SVM using Linear Kernel classifiers are applied for getting validation through experiments. These experimental outputs show the proposed oversampling method is superior giving accurate results in imbalanced data than the random oversampling approach.