Enhanced Support Vectors based Oversampling Method

Class distribution imbalance in the dataset is a critical factor in machine learning based classification performance. Many oversampling techniques were used to overcome the effect of this problem and retrieve the balance state for effective modelling. This paper presents an enhanced technique based on the Support Vectors for oversampling the minority class by generating synthetic data for better performance. The proposed method is tested on machine-generated datasets for binary classification problems. Experiments show an average improvement of approximately 15% in the classification recall and AUC performance measures on both datasets using the proposed method over the leading of the studied oversampling techniques. The proposed method can benefit the real-life machine learning classification problems where the positive class forms the minority of the training dataset like fraud-detection, medical diagnosis, and other problems, by expanding the positive class samples with more discriminative and evenly distributed related synthetic samples.