Received: 31 July 2019 Revised: 5 October 2019 Accepted: 25 October 2019
DOI: 10.1002/ett.3820
SPECIAL ISSUE ARTICLE
Sparse support vector machine with pinball loss
M. Tanveer
1
S. Sharma
1
R. Rastogi
2
P. Anand
2
1
Discipline of Mathematics, Indian
Institute of Technology Indore, Simrol,
Indore, India
2
Department of Computer Science, Faculty
of Mathematics and Computer Science,
South Asian University, Delhi, India
Correspondence
M. Tanveer, Discipline of Mathematics,
Indian Institute of Technology Indore,
Simrol, Indore, 453552, India.
Email: mtanveer@iiti.ac.in
Funding information
Early Career Research Award Scheme,
Grant/Award Number: ECR/2017/000053;
Research (EMR) Scheme, Grant/Award
Number: 22(0751)/17/EMR-II; Indian
Institute of Technology Indore
Abstract
The standard support vector machine (SVM) with a hinge loss function suffers
from feature noise sensitivity and instability. Employing a pinball loss function
instead of a hinge loss function in SVMs provides noise insensitivity to the model
as it maximizes the quantile distance. However, the pinball loss function simul-
taneously causes the model to lose sparsity by penalizing correctly classified
samples. To overcome the aforementioned shortcomings, we propose a novel
sparse SVM with pinball loss (Pin-SSVM) for solving classification problems.
The proposed Pin-SSVM employs L1-norm in the SVM classifier with the pinball
loss (Pin-SVM), which ensures the robustness, sparseness, and noise insensi-
tivity of the model. The proposed Pin-SSVM eradicates the need to solve the
dual as we simply obtain its solution by solving a linear programming problem
(LPP). The proposed Pin-SSVM does not spend more computational time as that
of Pin-SVM. Hence, solving an LPP with two linear inequality constraints does
not affect the computational complexity. The numerical experiments on several
real-world benchmark noise corrupted and imbalanced UCI datasets demon-
strate that the proposed Pin-SSVM is suitable for noisy and imbalanced data sets,
and in most cases, outperforms the results of the baseline models.
1 INTRODUCTION
Support vector machines (SVMs) are introduced to machine learning community by Vapnik and other researchers for clas-
sification and regression problems.
1-6
SVM has earned a great popularity as it has made substantial contribution to real-life
applications such as face detection,
7
cancer diagnosis,
8
electroencephalogram signal classification,
9
text categorization,
10
feature extraction
11
and biomedicine.
12
SVM has excellently performed on a wide variety of problems
7,13
as it constructs
a separating hyperplane such that the margin between the two supporting parallel hyperplanes is maximized. This idea
leads to an introduction of regularization term. Finally, an optimal separating hyperplane is selected to be the plane lying
in the middle of supporting hyperplanes. It incorporates the structural risk minimization principle of statistical learn-
ing theory.
14
Thenceforth, various models have been proposed for multiclass classification, including binary decision tree
SVM,
15
kernel method,
16
discriminative methods,
17
triclass SVM,
18
binary tree architecture,
19
decision directed acyclic
graph technique,
20
and multiclass pattern recognition.
21
The most challenging part in SVM is its high computational complexity as solving a large quadratic programming
problem (QPP) causes SVM to take longer training time. Later on, Fung and Mangasarian
22
introduced a Newton method
for feature selection for linear programming SVM (NLPSVM) which employs 1-norm in SVM formulation that is known
to generate a very sparse solution. NLPSVM suppresses the input feature space, which means that the separating hyper-
plane depends on a very few input vectors. Afterwards, various solvers have been proposed for obtaining solution of SVM
including SVMlight
13
and sequential minimal optimization.
23,24
Subsequently, Mangasarian
25
proposed an exact 1-norm
SVM, which is formulated as an unconstrained minimization of a convex differentiable piecewise-quadratic objective
Trans Emerging Tel Tech. 2020;e3820. wileyonlinelibrary.com/journal/ett © 2020 John Wiley & Sons, Ltd. 1 of 13
https://doi.org/10.1002/ett.3820