One-vs-all for class imbalance learning

The performance of support vector machines (SVMs) can deteriorate when the number of samples in one class is much greater than that in the other. Existing methods tackle this problem by modifying the learning algorithms or resampling the datasets. In this paper, we propose a new method called one-vs-all for class imbalance learning (OVACIL) which neither modifies the SVM learning algorithms nor resamples the datasets. In the OVACIL method, we re-group a given imbalanced dataset into a number of new datasets comprising of all the original samples and train standard SVM classifiers using each of the datasets. The output scores of these classifiers on a testing sample are then compared and a final decision is made without a fixed decision threshold. This comparison is not biased toward any particular class, resulting in high accuracies of both classes. The Gmean and Fmeasure values obtained by OVACIL on 18 real-world imbalanced datasets surpass the previous best values reported by other state-of-the-art CIL methods on most of these datasets.