Improved adaptive semi-unsupervised weighted oversampling using sparsity factor for imbalanced datasets

Abstract

With the incredible surge in data volumes, problems associated with data analysis have been increasingly complicated. In data mining algorithms, imbalanced data is a profound problem in machine learning paradigm. It appears due to desperate nature of data in which, one class with a large number of instances presents the majority class, while the other class with only a few instances is known as minority class. The classifier model biases towards the majority class and neglects the minority class which may happen to be the most essential class; resulting into costly misclassification error of minority class in real-world scenarios. Imbalanced data problem is significantly overcome by using re-sampling techniques, in which oversampling techniques are proven to be more effective than undersampling. This study proposes an Improved Adaptive Semi Unsupervised Weighted Oversampling (IA-SUWO) technique with sparsity factor, which efficiently solves between-the-class and within-the-class imbalances problem. Along with avoiding over-generalization, overfitting problems and removing noise from the data, this technique enhances the number of synthetic instances in the minority sub-clusters appropriately. A comprehensive experimental setup is used to evaluate the performance of the proposed approach. The comparative analysis reveals that the IA-SUWO performs better than the existing baseline oversampling techniques.