Rough margin based core vector machine

The recently proposed rough margin based support vector machine (RMSVM) could tackle the overfitting problem due to outliers effectively with the help of rough margins. However, the standard solvers for them are time consuming and not feasible for large datasets. On the other hand, the core vector machine (CVM) is an optimization technique based on the minimum enclosing ball that can scale up an SVM to handle very large datasets. While the 2-norm error used in the CVM might make it theoretically less robust against outliers, the rough margin could make up this deficiency. Therefore we propose our rough margin based core vector machine algorithms. Experimental results show that our algorithms hold the generalization performance almost as good as the RMSVM on large scale datasets and improve the accuracy of the CVM significantly on extremely noisy datasets, whilst cost much less computational resources and are often faster than the CVM.