Two effective sample selection methods for support vector machine

Article type: Research Article

Authors: Pan, Xianli | Xu, Yitian^*

Affiliations: College of Science, China Agricultural University, Beijing, China

Correspondence: [*] Corresponding author. Yitian Xu, College of Science, China Agricultural University, Beijing 100083, China. Tel.: +86 10 62737077; Fax: +86 10 62736563; E-mail: [email protected].

Abstract: The support vector machine (SVM) has gained prominence in classification problems. One characteristic of it is that the optimal decision hyperplane is only determined by support vectors (SVs). Inspired by this, we propose two novel sample selection methods based on the idea of potential SVs. One is the distance-based sample selection method (DSSM), it calculates the distance between each pair of samples from different classes, and then selects those nearer to the other class as potential SVs. The other is the K-nearest neighbors (Knn)-based sample selection method (KSSM), which is presented on account of the boundary nearest SVM (BNSVM). The BNSVM regards the nearest neighbors between two classes as potential SVs. While our KSSM extends the neighbor number from one in BNSVM to k. In addition, we present a method to choose an appropriate k. Afterwards, we do classification with these potential SVs. These selected samples contain useful information for classification which ensures the accuracy of classifiers. Moreover, the DSSM and KSSM work fast since they are only related to the distance and do not need to do iterations. Experiments conducted on small datasets and large-scaled ones show the validity of our DSSM and KSSM.

Keywords: SVM, Sample selection, DSSM, KSSM

DOI: 10.3233/IFS-151785

Journal: Journal of Intelligent & Fuzzy Systems, vol. 30, no. 2, pp. 659-670, 2016