Finding the samples near the decision plane for support vector learning

The decision plane of support vector machine (SVM) is decided by a few support vectors (SVs) in the training set. If there exist overlapping regions among different classes, SVs mainly locate in the overlap regions. A number of approaches have been proposed to find the samples in overlapping regions to condense the training set. However, the performance of these approaches would degrade if there is no overlapping region in the training set. In this paper, the extended nearest neighbor chain is proposed to find samples near the decision plane to avoid degrading performance in the cases of no overlapping region between different classes. Experimental results demonstrate that the proposed method performs better than the previous ones on artificial synthetic datasets as well as benchmark datasets. Additionally, the proposed method can obtain a higher compression ratio than previous ones.