| Abstract: |
| Imbalanced data classification problems appear quite commonly in real-world applications and impose great challenges to traditional classification approaches which work well only on balanced data but usually perform poorly on the minority class when the data is imbalanced. Resampling preprocessing by oversampling the minority class or downsampling the majority class helps improve the performance but may suffer from overfitting or loss of information. In this paper we propose a novel method called pairwise robust support vector machine (PRSVM) to overcome the difficulty of imbalanced data classification. It adapts the non-convex robust support vector classification loss to the pairwise learning setting. In the training process, samples from the minority class and the majority class always appear as pairs. This automatically balances the impact of two classes. Simulations and real-world applications show that PRSVM is highly effective. |
|