A new segmented oversampling method for imbalanced data classification using quasi-linear SVM

Data imbalance occurs on most real-world classification problems and decreases the performance of classifiers. An oversampling method addresses the imbalance problem by generating synthetic samples to balance the data distribution. However, many of the existing oversampling methods have potential problems in generating wrong and unnecessary synthetic samples, which makes the learning tasks difficult. This paper proposes a new segmented oversampling method for imbalanced data classification. The input space is first partitioned into several linearly separable local partitions along the potential separation boundary by introducing a bottom-up, minimal-spanning-tree-based clustering method; an oversampling method is then applied within each local linear partition to prevent the generation of wrong and unnecessary synthetic samples; a quasi-linear support vector machine is finally used to realize the classification by taking advantages of the local linear partitions. Simulation results on different real-world datasets show that the proposed segmented oversampling method is effective for imbalanced data classifications.