Union nonparallel support vector machines framework with consistency

doi:10.1016/j.asoc.2023.110129

Union nonparallel support vector machines framework with consistency

2023-03-01

2

PRE

AI

李春娜 (Chun‐Na Li)

邵

邵元海 (Yuan‐Hai Shao) *

DOI：10.1016/j.asoc.2023.110129

Original

Original request for help

Save

Abstract

En 中文

Though there are several dozens of nonparallel support vector machines (NSVMs), little studies on general forms and characteristics of NSVMs are investigated. To fill in this gap, this paper categorizes the existing NSVMs into two types and reveals the advantages and defects of different types of them. In particular, inconsistency problems for these models are pointed out and discussed. Based on this observation, this paper further proposes and investigates a novel max-min distance-based nonparallel support vector machine (NSVM) with desired consistency. Compared with the existing methods, the proposed NSVM has the consistency of training and test and the consistency of metric. In addition, NSVM also assigns each sample an ascertained loss, which not only is able to identify if a data sample being classified correctly, but also makes NSVM completely in line with its decision rule. NSVM can be easily extended to its nonlinear version, and both of them are effectively solved through a modified proximal difference-of-convex algorithm with extrapolation algorithm. Experimental results on the artificial dataset, benchmark datasets and a real-world dataset support the advantages of the proposed model.(c) 2023 Elsevier B.V. All rights reserved.

Keywords:

Support vector machines

Nonparallel support vector machines

Distance-based classifier

Multiclass classification

Consistency

Union nonparallel support vector machines framework with consistency

Abstract

Journal

Organization