As a kernel-based method, whether the selected kernel matches the data determines the performance of support vector machine. Conventional support vector classifiers are not suitable to the imbalanced learning tasks since they tend to classify the instances to the majority class which is the less important class. In this paper, we propose a weighted maximum margin criterion to optimize the data-dependent kernel, which makes the minority class more clustered in the induced feature space. We train support vector classification with the optimal kernel. The experimental results on nine benchmark data sets indicate the effectiveness of the proposed algorithm for imbalanced data classification problems.
Mathematical and Computer Modelling 08/2011; 54(3-4):1093-1099. DOI:10.1016/j.mcm.2010.11.040 · 2.02 Impact Factor