A Tutorial on -Support Vector Machines

Pai-Hsuen Chen
1
, Chih-Jen Lin
1
, and Bernhard Sch olkopf
2
1
Department of Computer Science and Information Engineering
National Taiwan University
Taipei 106, Taiwan
2
Max Planck Institute for Biological Cybernetics, T ubingen, Germany
[email protected]
Abstract. We briey describe the main ideas of statistical learning theory, sup-
port vector machines (SVMs), and kernel feature spaces. We place particular em-
phasis on a description of the so-called -SVM, including details of the algorithm
and its implementation, theoretical results, and practical applications.
1 An Introductory Example
Suppose we are given empirical data
(x
1
, y
1
), . . . , (x
m
, y
m
) A 1. (1)
Here, the domain A is some nonempty set that the patterns x
i
are taken from; the y
i
are called labels or targets.
Unless stated otherwise, indices i and j will always be understood to run over the
training set, i.e., i, j = 1, . . . , m.
Note that we have not made any assumptions on the domain A other than it being a
set. In order to study the problem of learning, we need additional structure. In learning,
we want to be able to generalize to unseen data points. In the case of pattern recognition,
this means that given some new pattern x A, we want to predict the corresponding
y 1. By this we mean, loosely speaking, that we choose y such that (x, y) is in
some sense similar to the training examples. To this end, we need similarity measures in
A and in 1. The latter is easy, as two target values can only be identical or different.
For the former, we require a similarity measure
k : A A R,
(x, x

) k(x, x

), (2)
i.e., a function that, given two examples x and x

, returns a real number characterizing

their similarity. For reasons that will become clear later, the function k is called a kernel
([24], [1], [8]).
A type of similarity measure that is of particular mathematical appeal are dot prod-
ucts. For instance, given two vectors x, x

R
N
, the canonical dot product is dened

Parts of the present article are based on [31].

as
(x x

) :=
N

i=1
(x)
i
(x

)
i
. (3)
Here, (x)
i
denotes the ith entry of x.
The geometrical interpretation of this dot product is that it computes the cosine of
the angle between the vectors x and x

, provided they are normalized to length 1. More-

over, it allows computation of the length of a vector x as
_
(x x), and of the distance
between two vectors as the length of the difference vector. Therefore, being able to
compute dot products amounts to being able to carry out all geometrical constructions
that can be formulated in terms of angles, lengths and distances.
Note, however, that we have not made the assumption that the patterns live in a
dot product space. In order to be able to use a dot product as a similarity measure, we
therefore rst need to transform them into some dot product space H, which need not
be identical to R
N
. To this end, we use a map
: A H
x x. (4)
The space H is called a feature space. To summarize, there are three benets to trans-
form the data into H
1. It lets us dene a similarity measure from the dot product in H,
k(x, x

) := (x x

) = ((x) (x

)). (5)
2. It allows us to deal with the patterns geometrically, and thus lets us study learning
algorithm using linear algebra and analytic geometry.
3. The freedom to choose the mapping will enable us to design a large variety of
learning algorithms. For instance, consider a situation where the inputs already live
in a dot product space. In that case, we could directly dene a similarity measure
as the dot product. However, we might still choose to rst apply a nonlinear map
to change the representation into one that is more suitable for a given problem and
learning algorithm.
We are now in the position to describe a pattern recognition learning algorithm that
is arguable one of the simplest possible. The basic idea is to compute the means of the
two classes in feature space,
c
+
=
1
m
+

{i:y
i
=+1}
x
i
, (6)
c

=
1
m

{i:y
i
=1}
x
i
, (7)
where m
+
and m

are the number of examples with positive and negative labels, re-
spectively (see Figure 1). We then assign a new point x to the class whose mean is
o
+
+
+
+
o
o
c
+
c
-
x-c
w
x
c
.
Fig. 1. A simple geometric classication algorithm: given two classes of points (depicted by o
and +), compute their means c
+
, c

and assign a test pattern x to the one whose mean is closer.

This can be done by looking at the dot product between x c (where c = (c
+
+ c

)/2) and
w := c
+
c

, which changes sign as the enclosed angle passes through /2. Note that the
corresponding decision boundary is a hyperplane (the dotted line) orthogonal to w (from [31]).
closer to it. This geometrical construction can be formulated in terms of dot products.
Half-way in between c
+
and c

lies the point c := (c

+
+c

)/2. We compute the class

of x by checking whether the vector connecting c and x encloses an angle smaller than
/2 with the vector w := c
+
c

connecting the class means, in other words

y = sgn ((x c) w)
y = sgn ((x (c
+
+ c

)/2) (c
+
c

))
= sgn ((x c
+
) (x c

) + b). (8)
Here, we have dened the offset
b :=
1
2
_
|c

|
2
|c
+
|
2
_
. (9)
It will be proved instructive to rewrite this expression in terms of the patterns x
i
in
the input domain A. To this end, note that we do not have a dot product in A, all we
have is the similarity measure k (cf. (5)). Therefore, we need to rewrite everything in
terms of the kernel k evaluated on input patterns. To this end, substitute (6) and (7) into
(8) to get the decision function
y = sgn
_
_
1
m
+

{i:y
i
=+1}
(x x
i
)
1
m

{i:y
i
=1}
(x x
i
) + b
_
_
= sgn
_
_
1
m
+

{i:y
i
=+1}
k(x, x
i
)
1
m

{i:y
i
=1}
k(x, x
i
) + b
_
_
. (10)
Similarly, the offset becomes
b :=
1
2
_
_
1
m
2

{(i,j):y
i
=y
j
=1}
k(x
i
, x
j
)
1
m
2
+

{(i,j):y
i
=y
j
=+1}
k(x
i
, x
j
)
_
_
. (11)
Let us consider one well-known special case of this type of classier. Assume that the
class means have the same distance to the origin (hence b = 0), and that k can be viewed
as a density, i.e., it is positive and has integral 1,
_
X
k(x, x

)dx = 1 for all x

A. (12)
In order to state this assumption, we have to require that we can dene an integral on
A.
If the above holds true, then (10) corresponds to the so-called Bayes decision bound-
ary separating the two classes, subject to the assumption that the two classes were gen-
erated from two probability distributions that are correctly estimated by the Parzen
windows estimators of the two classes,
p
1
(x) :=
1
m
+

{i:y
i
=+1}
k(x, x
i
) (13)
p
2
(x) :=
1
m

{i:y
i
=1}
k(x, x
i
). (14)
Given some point x, the label is then simply computed by checking which of the two,
p
1
(x) or p
2
(x), is larger, which directly leads to (10). Note that this decision is the best
we can do if we have no prior information about the probabilities of the two classes.
For further details, see [31].
The classier (10) is quite close to the types of learning machines that we will
be interested in. It is linear in the feature space, and while in the input domain, it is
represented by a kernel expansion in terms of the training points. It is example-based
in the sense that the kernels are centered on the training examples, i.e., one of the two
arguments of the kernels is always a training example. The main points that the more
sophisticated techniques to be discussed later will deviate from (10) are in the selection
of the examples that the kernels are centered on, and in the weights that are put on the
individual data in the decision function. Namely, it will no longer be the case that all
training examples appear in the kernel expansion, and the weights of the kernels in the
expansion will no longer be uniform. In the feature space representation, this statement
corresponds to saying that we will study all normal vectors w of decision hyperplanes
that can be represented as linear combinations of the training examples. For instance,
we might want to remove the inuence of patterns that are very far away from the
decision boundary, either since we expect that they will not improve the generalization
error of the decision function, or since we would like to reduce the computational cost
of evaluating the decision function (cf. (10)). The hyperplane will then only depend on
a subset of training examples, called support vectors.
2 Learning Pattern Recognition from Examples
With the above example in mind, let us now consider the problem of pattern recognition
in a more formal setting ([37], [38]), following the introduction of [30]. In two-class
pattern recognition, we seek to estimate a function
f : A 1 (15)
based on input-output training data (1). We assume that the data were generated inde-
pendently from some unknown (but xed) probability distribution P(x, y). Our goal
is to learn a function that will correctly classify unseen examples (x, y), i.e., we want
f(x) = y for examples (x, y) that were also generated from P(x, y).
If we put no restriction on the class of functions that we choose our estimate f
from, however, even a function which does well on the training data, e.g. by satisfying
f(x
i
) = y
i
for all i = 1, . . . , m, need not generalize well to unseen examples. To see
this, note that for each function f and any test set ( x
1
, y
1
), . . . , ( x
m
, y
m
) R
N
1,
satisfying x
1
, . . . , x
m
x
1
, . . . , x
m
= , there exists another function f

such
that f

(x
i
) = f(x
i
) for all i = 1, . . . , m, yet f

( x
i
) ,= f( x
i
) for all i = 1, . . . , m.
As we are only given the training data, we have no means of selecting which of the two
functions (and hence which of the completely different sets of test label predictions) is
preferable. Hence, only minimizing the training error (or empirical risk),
R
emp
[f] =
1
m
m

i=1
1
2
[f(x
i
) y
i
[, (16)
does not imply a small test error (called risk), averaged over test examples drawn from
the underlying distribution P(x, y),
R[f] =
_
1
2
[f(x) y[ dP(x, y). (17)
Statistical learning theory ([41], [37], [38], [39]), or VC (Vapnik-Chervonenkis) theory,
shows that it is imperative to restrict the class of functions that f is chosen from to one
which has a capacity that is suitable for the amount of available training data. VC theory
provides bounds on the test error. The minimization of these bounds, which depend on
both the empirical risk and the capacity of the function class, leads to the principle of
structural risk minimization ([37]). The best-known capacity concept of VC theory is
the VC dimension, dened as the largest number h of points that can be separated in
all possible ways using functions of the given class. An example of a VC bound is the
following: if h < m is the VC dimension of the class of functions that the learning
machine can implement, then for all functions of that class, with a probability of at
least 1 , the bound
R(f) R
emp
(f) +
_
h
m
,
log()
m
_
(18)
holds, where the condence term is dened as

_
h
m
,
log()
m
_
=

h
_
log
2m
h
+ 1
_
log(/4)
m
. (19)
Tighter bounds can be formulated in terms of other concepts, such as the annealed VC
entropy or the Growth function. These are usually considered to be harder to evaluate,
but they play a fundamental role in the conceptual part of VC theory ([38]). Alterna-
tive capacity concepts that can be used to formulate bounds include the fat shattering
dimension ([2]).
The bound (18) deserves some further explanatory remarks. Suppose we wanted to
learn a dependency where P(x, y) = P(x) P(y), i.e., where the pattern x contains
no information about the label y, with uniform P(y). Given a training sample of xed
size, we can then surely come up with a learning machine which achieves zero training
error (provided we have no examples contradicting each other). However, in order to
reproduce the random labelling, this machine will necessarily require a large VC di-
mension h. Thus, the condence term (19), increasing monotonically with h, will be
large, and the bound (18) will not support possible hopes that due to the small training
error, we should expect a small test error. This makes it understandable how (18) can
hold independent of assumptions about the underlying distribution P(x, y): it always
holds (provided that h < m), but it does not always make a nontrivial prediction a
bound on an error rate becomes void if it is larger than the maximum error rate. In order
to get nontrivial predictions from (18), the function space must be restricted such that
the capacity (e.g. VC dimension) is small enough (in relation to the available amount
of data).
3 Hyperplane Classiers
In the present section, we shall describe a hyperplane learning algorithm that can be
performed in a dot product space (such as the feature space that we introduced previ-
ously). As described in the previous section, to design learning algorithms, one needs
to come up with a class of functions whose capacity can be computed.
[42] considered the class of hyperplanes
(w x) + b = 0 w R
N
, b R, (20)
corresponding to decision functions
f(x) = sgn ((w x) + b), (21)
and proposed a learning algorithm for separable problems, termed the Generalized Por-
trait, for constructing f from empirical data. It is based on two facts. First, among all
hyperplanes separating the data, there exists a unique one yielding the maximummargin
of separation between the classes,
max
w,b
min|x x
i
| : x R
N
, (w x) + b = 0, i = 1, . . . , m. (22)
Second, the capacity decreases with increasing margin.
To construct this Optimal Hyperplane (cf. Figure 2), one solves the following opti-
mization problem:
minimize
w,b
1
2
|w|
2
subject to y
i
((w x
i
) + b) 1, i = 1, . . . , m. (23)
.
w
{x | (w x) + b = 0}
.
{x | (w x) + b = 1}
.
{x | (w x) + b = +1}
.
x
2
x
1
Note:
(w x
1
) + b = +1
(w x
2
) + b = 1
=> (w (x
1
x
2
)) = 2
=> (x
1
x
2
) =
w
||w||
(
)
.
.
.
.
2
||w||
y
i
= 1
y
i
= +1 r
r
r
r
r
x
x
x
x
Fig. 2. Abinary classication toy problem: separate balls fromdiamonds. The optimal hyperplane
is orthogonal to the shortest line connecting the convex hulls of the two classes (dotted), and
intersects it half-way between the two classes. The problem is separable, so there exists a weight
vector w and a threshold b such that y
i
((w x
i
) + b) > 0 (i = 1, . . . , m). Rescaling w and b
such that the point(s) closest to the hyperplane satisfy |(w x
i
) + b| = 1, we obtain a canonical
form (w, b) of the hyperplane, satisfying y
i
((w x
i
)+b) 1. Note that in this case, the margin,
measured perpendicularly to the hyperplane, equals 2/w. This can be seen by considering two
points x
1
, x
2
on opposite sides of the margin, i.e., (w x
1
) + b = 1, (w x
2
) + b = 1, and
projecting them onto the hyperplane normal vector w/w (from [29]).
A way to solve (23) is through its Lagrangian dual:
max
0
(min
w,b
L(w, b, )), (24)
where
L(w, b, ) =
1
2
|w|
2

i=1

i
(y
i
((x
i
w) + b) 1) . (25)
The Lagrangian L has to be minimized with respect to the primal variables w and
b and maximized with respect to the dual variables
i
. For a nonlinear problem like
(23), called the primal problem, there are several closely related problems of which
the Lagrangian dual is an important one. Under certain conditions, the primal and dual
problems have the same optimal objective values. Therefore, we can instead solve the
dual which may be an easier problem than the primal. In particular, we will see in
Section 4 that when working in feature spaces, solving the dual may be the only way to
train SVM.
Let us try to get some intuition for this primal-dual relation. Assume ( w,

b) is an
optimal solution of the primal with the optimal objective value =
1
2
| w|
2
. Thus, no
(w, b) satises
1
2
|w|
2
< and y
i
((w x
i
) + b) 1, i = 1, . . . , m. (26)
With (26), there is 0 such that for all w, b
1
2
|w|
2

m

i=1

i
(y
i
((x
i
w) + b) 1) 0. (27)
We do not provide a rigorous proof here but details can be found in, for example, [5].
Note that for general convex programming this result requires some additional condi-
tions on constraints which are now satised by our simple linear inequalities.
Therefore, (27) implies
max
0
min
w,b
L(w, b, ) . (28)
On the other hand, for any ,
min
w,b
L(w, b, ) L( w,

b, ),
so
max
0
min
w,b
L(w, b, ) max
0
L( w,

b, ) =
1
2
| w|
2
= . (29)
Therefore, with (28), the inequality in (29) becomes an equality. This property is
the strong duality where the primal and dual have the same optimal objective value. In
addition, putting ( w,

b) into (27), with

i
0 and y
i
((x
i
w) +

b) 1 0,

i
[y
i
((x
i
w) +

b) 1] = 0, i = 1, . . . , m, (30)
which is usually called the complementarity condition.
To simplify the dual, as L(w, b, ) is convex when is xed, for any given ,

b
L(w, b, ) = 0,

w
L(w, b, ) = 0, (31)
leads to
m

i=1

i
y
i
= 0 (32)
and
w =
m

i=1

i
y
i
x
i
. (33)
As is now given, we may wonder what (32) means. From the denition of the La-
grangian, if

m
i=1

i
y
i
,= 0, we can decrease b

m
i=1

i
y
i
in L(w, b, ) as much as
we want. Therefore, by substituting (33) into (24), the dual problem can be written as
max
0
_

m
i=1

i

1
2

m
i,j=1

i

j
y
i
y
j
(x
i
x
j
) if

m
i=1

i
y
i
= 0,
if

m
i=1

i
y
i
,= 0.
(34)
As is denitely not the maximal objective value of the dual, the dual optimal so-
lution does not happen when

m
i=1

i
y
i
,= 0. Therefore, the dual problem is simplied
to nding multipliers
i
which
maximize
R
m
m

i=1

i

1
2
m

i,j=1

j
y
i
y
j
(x
i
x
j
) (35)
subject to
i
0, i = 1, . . . , m, and
m

i=1

i
y
i
= 0. (36)
This is the dual SVM problem that we usually refer to. Note that (30), (32),
i
0i,
and (33), are called the Karush-Kuhn-Tucker (KKT) optimality conditions of the primal
problem. Except an abnormal situation where all optimal
i
are zero, b can be computed
using (30).
The discussion from (31) to (33) implies that we can consider a different form of
dual problem:
maximize
w,b,0
L(w, b, )
subject to

b
L(w, b, ) = 0,

w
L(w, b, ) = 0.
(37)
This is the so called Wolfe dual for convex optimization, which is a very early work in
duality [45]. For convex and differentiable problems, it is equivalent to the Lagrangian
dual though the derivation of the Lagrangian dual more easily shows the strong duality
results. Some notes about the two duals are in, for example, [3, Section 5.4].
Following the above discussion, the hyperplane decision function can be written as
f(x) = sgn
_
m

i=1
y
i

i
(x x
i
) + b
_
. (38)
The solution vector w thus has an expansion in terms of a subset of the training pat-
terns, namely those patterns whose
i
is non-zero, called Support Vectors. By (30), the
Support Vectors lie on the margin (cf. Figure 2). All remaining examples of the training
set are irrelevant: their constraint (23) does not play a role in the optimization, and they
do not appear in the expansion (33). This nicely captures our intuition of the problem:
as the hyperplane (cf. Figure 2) is completely determined by the patterns closest to it,
the solution should not depend on the other examples.
The structure of the optimization problem closely resembles those that typically
arise in Lagranges formulation of mechanics. Also there, often only a subset of the
constraints become active. For instance, if we keep a ball in a box, then it will typically
roll into one of the corners. The constraints corresponding to the walls which are not
touched by the ball are irrelevant, the walls could just as well be removed.
Seen in this light, it is not too surprising that it is possible to give a mechanical in-
terpretation of optimal margin hyperplanes ([9]): If we assume that each support vector
x
i
exerts a perpendicular force of size
i
and sign y
i
on a solid plane sheet lying along
the hyperplane, then the solution satises the requirements of mechanical stability. The
constraint (32) states that the forces on the sheet sum to zero; and (33) implies that the
torques also sum to zero, via

i
x
i
y
i

i
w/|w| = w w/|w| = 0.
There are theoretical arguments supporting the good generalization performance of
the optimal hyperplane ([41], [37], [4], [33], [44]). In addition, it is computationally
attractive, since it can be constructed by solving a quadratic programming problem.
4 Optimal Margin Support Vector Classiers
We now have all the tools to describe support vector machines ([38], [31]). Everything
in the last section was formulated in a dot product space. We think of this space as the
feature space H described in Section 1. To express the formulas in terms of the input
patterns living in A, we thus need to employ (5), which expresses the dot product of
bold face feature vectors x, x

in terms of the kernel k evaluated on input patterns x, x

,
k(x, x

) = (x x

). (39)
This can be done since all feature vectors only occurred in dot products. The weight
vector (cf. (33)) then becomes an expansion in feature space,
1
and will thus typically
no longer correspond to the image of a single vector from input space. We thus obtain
decision functions of the more general form (cf. (38))
f(x) = sgn
_
m

i=1
y
i

i
((x) (x
i
)) + b
_
= sgn
_
m

i=1
y
i

i
k(x, x
i
) + b
_
, (40)
and the following quadratic program (cf. (35)):
maximize
R
m
W() =
m

i=1

i

1
2
m

i,j=1

j
y
i
y
j
k(x
i
, x
j
) (41)
subject to
i
0, i = 1, . . . , m, and
m

i=1

i
y
i
= 0. (42)
Working in the feature space somewhat forces us to solve the dual problem instead
of the primal. The dual problem has the same number of variables as the number of
training data. However, the primal problem may have a lot more (even innite) variables
depending on the dimensionality of the feature space (i.e. the length of (x)). Though
our derivation of the dual problem in Section 3 considers problems in nite-dimensional
spaces, it can be directly extended to problems in Hilbert spaces [20].
1
This constitutes a special case of the so-called representer theorem, which states that under
fairly general conditions, the minimizers of objective functions which contain a penalizer in
terms of a norm in feature space will have kernel expansions ([43], [31]).
Fig. 3. Example of a Support Vector classier found by using a radial basis function kernel
k(x, x

) = exp(x x

2
). Both coordinate axes range from -1 to +1. Circles and disks
are two classes of training examples; the middle line is the decision surface; the outer lines pre-
cisely meet the constraint (23). Note that the Support Vectors found by the algorithm (marked by
extra circles) are not centers of clusters, but examples which are critical for the given classica-
tion task. Grey values code the modulus of the argument
P
m
i=1
y
i

i
k(x, x
i
) +b of the decision
function (40) (from [29]).)
5 Kernels
We now take a closer look at the issue of the similarity measure, or kernel, k. In this
section, we think of A as a subset of the vector space R
N
, (N N), endowed with the
canonical dot product (3).
5.1 Product Features
Suppose we are given patterns x R
N
where most information is contained in the dth
order products (monomials) of entries [x]
j
of x,
[x]
j
1
[x]
j
d
, (43)
where j
1
, . . . , j
d
1, . . . , N. In that case, we might prefer to extract these product
features, and work in the feature space H of all products of d entries. In visual recog-
nition problems, where images are often represented as vectors, this would amount to
extracting features which are products of individual pixels.
For instance, in R
2
, we can collect all monomial feature extractors of degree 2 in
the nonlinear map
: R
2
H = R
3
(44)
([x]
1
, [x]
2
) ([x]
2
1
, [x]
2
2
, [x]
1
[x]
2
). (45)
This approach works ne for small toy examples, but it fails for realistically sized prob-
lems: for N-dimensional input patterns, there exist
N
H
=
(N + d 1)!
d!(N 1)!
(46)
different monomials (43), comprising a feature space H of dimensionality N
H
. For
instance, already 16 16 pixel input images and a monomial degree d = 5 yield a
dimensionality of 10
10
.
In certain cases described below, there exists, however, a way of computing dot
products in these high-dimensional feature spaces without explicitly mapping into them:
by means of kernels nonlinear in the input space R
N
. Thus, if the subsequent process-
ing can be carried out using dot products exclusively, we are able to deal with the high
dimensionality.
5.2 Polynomial Feature Spaces Induced by Kernels
In order to compute dot products of the form ((x) (x

)), we employ kernel repre-

sentations of the form
k(x, x

) = ((x) (x

)), (47)
which allow us to compute the value of the dot product in H without having to carry
out the map . This method was used by Boser et al. to extend the Generalized Por-
trait hyperplane classier [41] to nonlinear Support Vector machines [8]. Aizerman et
al. called H the linearization space, and used in the context of the potential function
classication method to express the dot product between elements of H in terms of
elements of the input space [1].
What does k look like for the case of polynomial features? We start by giving an
example ([38]) for N = d = 2. For the map

2
: ([x]
1
, [x]
2
) ([x]
2
1
, [x]
2
2
, [x]
1
[x]
2
, [x]
2
[x]
1
), (48)
dot products in H take the form
(
2
(x)
2
(x

)) = [x]
2
1
[x

]
2
1
+ [x]
2
2
[x

]
2
2
+ 2[x]
1
[x]
2
[x

]
1
[x

]
2
= (x x

)
2
, (49)
i.e., the desired kernel k is simply the square of the dot product in input space. Note that
it is possible to modify (x x

)
d
such that it maps into the space of all monomials up to
degree d, dening ([38])
k(x, x

) = ((x x

) + 1)
d
. (50)
5.3 Examples of Kernels
When considering feature maps, it is also possible to look at things the other way
around, and start with the kernel. Given a kernel function satisfying a mathematical
condition termed positive deniteness, it is possible to construct a feature space such
that the kernel computes the dot product in that feature space. This has been brought
to the attention of the machine learning community by [1], [8], and [38]. In functional
analysis, the issue has been studied under the heading of Reproducing kernel Hilbert
space (RKHS).
Besides (50), a popular choice of kernel is the Gaussian radial basis function ([1])
k(x, x

) = exp
_
|x x

|
2
_
. (51)
An illustration is in Figure 3. For an overview of other kernels, see [31].
6 -Soft Margin Support Vector Classiers
In practice, a separating hyperplane may not exist, e.g. if a high noise level causes a
large overlap of the classes. To allow for the possibility of examples violating (23), one
introduces slack variables ([15], [38], [32])

i
0, i = 1, . . . , m (52)
in order to relax the constraints to
y
i
((w x
i
) + b) 1
i
, i = 1, . . . , m. (53)
A classier which generalizes well is then found by controlling both the classier ca-
pacity (via |w|) and the sum of the slacks

i

i
. The latter is done as it can be shown
to provide an upper bound on the number of training errors which leads to a convex
optimization problem.
One possible realization,called C-SVC, of a soft margin classier is minimizing the
objective function
(w, ) =
1
2
|w|
2
+ C
m

i=1

i
(54)
subject to the constraints (52) and (53), for some value of the constant C > 0 deter-
mining the trade-off. Here and below, we use boldface Greek letters as a shorthand
for corresponding vectors = (
1
, . . . ,
m
). Incorporating kernels, and rewriting it
in terms of Lagrange multipliers, this again leads to the problem of maximizing (41),
subject to the constraints
0
i
C, i = 1, . . . , m, and
m

i=1

i
y
i
= 0. (55)
The only difference from the separable case is the upper bound C on the Lagrange mul-
tipliers
i
. This way, the inuence of the individual patterns (which could be outliers)
gets limited. As above, the solution takes the form (40).
Another possible realization,called -SVC of a soft margin variant of the optimal
hyperplane uses the -parameterization ([32]). In it, the parameter C is replaced by a
parameter [0, 1] which is the lower and upper bound on the number of examples
that are support vectors and that lie on the wrong side of the hyperplane, respectively.
As a primal problem for this approach, termed the -SV classier, we consider
minimize
wH,R
m
,,bR
(w, , ) =
1
2
|w|
2
+
1
m
m

i=1

i
(56)
subject to y
i
(x
i
, w) + b)
i
, i = 1, . . . , m (57)
and
i
0, 0. (58)
Note that no constant C appears in this formulation; instead, there is a parameter , and
also an additional variable to be optimized. To understand the role of , note that for
= 0, the constraint (57) simply states that the two classes are separated by the margin
2/|w|.
To explain the signicance of , let us rst introduce the term margin error: by this,
we denote training points with
i
> 0. These are points which either are errors, or lie
within the margin. Formally, the fraction of margin errors is
R

emp
[g] :=
1
m
[i[y
i
g(x
i
) < [ . (59)
Here, g is used to denote the argument of the sign in the decision function (40): f =
sgn g. We are now in a position to state a result that explains the signicance of .
Proposition 1 ([32]). Suppose we run -SVC with kernel function k on some data with
the result that > 0. Then
(i) is an upper bound on the fraction of margin errors (and hence also on the fraction
of training errors).
(ii) is a lower bound on the fraction of SVs.
(iii) Suppose the data (x
1
, y
1
), . . . , (x
m
, y
m
) were generated iid from a distribution
Pr(x, y) = Pr(x) Pr(y[x), such that neither Pr(x, y = 1) nor Pr(x, y = 1) con-
tains any discrete component. Suppose, moreover, that the kernel used is analytic
and non-constant. With probability 1, asymptotically, equals both the fraction of
SVs and the fraction of margin errors.
Before we get into the technical details of the dual derivation, let us take a look at
a toy example illustrating the inuence of (Figure 4). The corresponding fractions of
SVs and margin errors are listed in table 1.
Let us next derive the dual of the -SV classication algorithm. We consider the
Lagrangian
L(w, , b, , , , ) =
1
2
|w|
2
+
1
m
m

i=1

i=1
(
i
(y
i
(x
i
, w) + b) +
i
) +
i

i
), (60)
Fig. 4. Toy problem (task: to separate circles from disks) solved using -SV classication, with
parameter values ranging from = 0.1 (top left) to = 0.8 (bottom right). The larger we make
, the more points are allowed to lie inside the margin (depicted by dotted lines). Results are
shown for a Gaussian kernel, k(x, x

) = exp(x x

2
) (from [31]).
Table 1. Fractions of errors and SVs, along with the margins of class separation, for the toy
example in Figure 4.
Note that upper bounds the fraction of errors and lower bounds the fraction of SVs, and that
increasing , i.e., allowing more errors, increases the margin.
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
fraction of errors 0.00 0.07 0.25 0.32 0.39 0.50 0.61 0.71
fraction of SVs 0.29 0.36 0.43 0.46 0.57 0.68 0.79 0.86
margin /w 0.005 0.018 0.115 0.156 0.364 0.419 0.461 0.546
using multipliers
i
,
i
, 0. This function has to be minimized with respect to the
primal variables w, , b, , and maximized with respect to the dual variables , , .
Following the same derivation in (31)(33), we compute the corresponding partial
derivatives and set them to 0, obtaining the following conditions:
w =
m

i=1

i
y
i
x
i
, (61)

i
+
i
= 1/m, (62)
m

i=1

i
y
i
= 0, (63)
m

i=1

i
= . (64)
Again, in the SV expansion (61), the
i
that are non-zero correspond to a constraint (57)
which is precisely met.
Substituting (61) and (62) into L, using
i
,
i
, 0, and incorporating kernels
for dot products, leaves us with the following quadratic optimization problem for -SV
classication:
maximize
R
m
W() =
1
2
m

i,j=1

j
y
i
y
j
k(x
i
, x
j
) (65)
subject to 0
i

1
m
, (66)
m

i=1

i
y
i
= 0, (67)
m

i=1

i
. (68)
As above, the resulting decision function can be shown to take the form
f(x) = sgn
_
m

i=1

i
y
i
k(x, x
i
) + b
_
. (69)
Compared with the C-SVC dual ((41), (55)), there are two differences. First, there is
an additional constraint (68). Second, the linear term

m
i=1

i
no longer appears in the
objective function (65). This has an interesting consequence: (65) is now quadratically
homogeneous in . It is straightforward to verify that the same decision function is
obtained if we start with the primal function
(w, , ) =
1
2
|w|
2
+ C
_
+
1
m
m

i=1

i
_
, (70)
i.e., if one does use C [31].
The computation of the threshold b and the margin parameter will be discussed in
Section 7.4.
A connection to standard SV classication, and a somewhat surprising interpreta-
tion of the regularization parameter C, is described by the following result:
Proposition 2 (Connection -SVC C-SVC [32]). If -SV classication leads to
> 0, then C-SV classication, with C set a priori to 1/m, leads to the same decision
function.
For further details on the connection between -SVMs and C-SVMs, see [16, 6]. By
considering the optimal as a function of parameters, a complete account is as follows:
Proposition 3 (Detailed connection -SVC C-SVC [11]).

m
i=1

i
/(Cm) by the
C-SVM is a well dened decreasing function of C. We can dene
lim
C

m
i=1

i
Cm
=
min
0 and lim
C0

m
i=1

i
Cm
=
max
1. (71)
Then,
1.
max
= 2 min(m
+
, m

)/m.
2. For any >
max
, the dual -SVM is infeasible. That is, the set of feasible points is
empty. For any (
min
,
max
], the optimal solution set of dual -SVM is the same
as that of either one or some C-SVM where these C form an interval. In addition,
the optimal objective value of -SVM is strictly positive. For any 0
min
,
dual -SVM is feasible with zero optimal objective value.
3. If the kernel matrix is positive denite, then
min
= 0.
Therefore, for a given problemand kernel, there is an interval [
min
,
max
] of admissible
values for , with 0
min

max
1. An illustration of the relation between and
C is in Figure 5.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
-3 -2 -1 0 1 2 3 4 5

log
10
C
Fig. 5. The relation between and C (using the RBF kernel on the problem australian from the
Statlog collection [25])
It has been noted that -SVMs have an interesting interpretation in terms of reduced
convex hulls [16, 6]. One can show that for separable problems, one can obtain the opti-
mal margin separating hyperplane by forming the convex hulls of both classes, nding
the shortest connection between the two convex hulls (since the problem is separable,
they are disjoint), and putting the hyperplane halfway along that connection, orthogonal
to it. If a problem is non-separable, however, the convex hulls of the two classes will
no longer be disjoint. Therefore, it no longer makes sense to search for the shortest line
connecting them. In this situation, it seems natural to reduce the convex hulls in size,
by limiting the size of the coefcients c
i
in the convex sets
C

:=
_

y
i
=1
c
i
x
i

y
i
=1
c
i
= 1, c
i
0
_
. (72)
to some value (0, 1). Intuitively, this amounts to limiting the inuence of individual
points. It is possible to show that the -SVM formulation solves the problem of nding
the hyperplane orthogonal to the closest line connecting the reduced convex hulls [16].
We now move on to another aspect of soft margin classication. When we intro-
duced the slack variables, we did not attempt to justify the fact that in the objective
function, we used a penalizer

m
i=1

i
. Why not use another penalizer, such as

m
i=1

p
i
,
for some p 0 [15]? For instance, p = 0 would yield a penalizer that exactly counts
the number of margin errors. Unfortunately, however, it is also a penalizer that leads to
a combinatorial optimization problem. Penalizers yielding optimization problems that
are particularly convenient, on the other hand, are obtained for p = 1 and p = 2. By
default, we use the former, as it possesses an additional property which is statistically
attractive. As the following proposition shows, linearity of the target function in the
slack variables
i
leads to a certain outlier resistance of the estimator. As above, we
use the shorthand x
i
for (x
i
).
Proposition 4 (Resistance of SV classication [32]). Suppose w can be expressed in
terms of the SVs which are not at bound,
w =
m

i=1

i
x
i
(73)
with
i
,= 0 only if
i
(0, 1/m) (where the
i
are the coefcients of the dual solu-
tion). Then local movements of any margin error x
j
parallel to w do not change the
hyperplane.
2
This result is about the stability of classiers. Results have also shown that in general
p = 1 leads to fewer support vectors. Further results in support of the p = 1 case can
be seen in [34, 36].
Although proposition 1 shows that possesses an intuitive meaning, it is still un-
clear how to choose for a learning task. [35] proves that given

R, a close upper bound
on the expected optimal Bayes risk, an asymptotically good estimate of the optimal
value of is 2

R:
Proposition 5. If R[f] is the expected risk dened in (17),
R
p
:= inf
f
R[f], (74)
and the kernel used by -SVM is universal, then for all > 2R
p
and all > 0, there
exists a constant c > 0 such that
P(T = (x
1
, y
1
), . . . , (x
m
, y
m
) [ R[f

T
] R
p
+ ) 1 e
cm
. (75)
Quite a few popular kernels such as the Gaussian are universal. The denition of a
universal kernel can be seen in [35]. Here, f

T
is the decision function obtained by
training -SVM on the data set T.
2
Note that the perturbation of the point is carried out in feature space. What it precisely corre-
sponds to in input space therefore depends on the specic kernel chosen.
Therefore, given an upper bound

R on R
p
, the decision function with respect to = 2

R
almost surely achieves a risk not larger than R
p
+ 2(

R R
p
).
The selection of and kernel parameters can be done by estimating the performance
of support vector binary classiers on data not yet observed. One such performance esti-
mate is the leave-one-out error, which is an almost unbiased estimate of the generaliza-
tion performance [22]. To compute this performance metric, a single point is excluded
from the training set, and the classier is trained using the remaining points. It is then
determined whether this new classier correctly labels the point that was excluded. The
process is repeated over the entire training set. Although theoretically attractive, this
estimate obviously entails a large computational cost.
Three estimates of the leave-one-out error for the -SV learning algorithm are pre-
sented in [17]. Of these three estimates, the general -SV bound is an upper bound on
the leave-one-out error, the restricted -SV estimate is an approximation that assumes
the sets of margin errors and support vectors on the margin to be constant, and the max-
imized target estimate is an approximation that assumes the sets of margin errors and
non-support vectors not to decrease. The derivation of the general -SV bound takes
a form similar to an upper bound described in [40] for the C-SV classier, while the
restricted -SV estimate is based on a similar C-SV estimate proposed in [40, 26]: both
these estimates are based on the geometrical concept of the span, which is (roughly
speaking) a measure of how easily a particular point in the training sample can be
replaced by the other points used to dene the classication function. No analogous
method exists in the C-SV case for the maximized target estimate.
7 Implementation of -SV Classiers
We change the dual form of -SV classiers to be a minimization problem:
minimize
R
m
W() =
1
2
m

i,j=1

j
y
i
y
j
k(x
i
, x
j
)
subject to 0
i

1
m
, (76)
m

i=1

i
y
i
= 0,
m

i=1

i
= . (77)
[11] proves that for any given , there is at least an optimal solution which satises
e
T
= . Therefore, it is sufcient to solve a simpler problem with the equality con-
straint (77).
Similar to C-SVC, the difculty of solving (76) is that y
i
y
j
k(x
i
, x
j
) are in gen-
eral not zero. Thus, for large data sets, the Hessian (second derivative) matrix of the
objective function cannot be stored in the computer memory, so traditional optimiza-
tion methods such as Newton or quasi Newton cannot be directly used. Currently, the
decomposition method is the most used approach to conquer this difculty. Here, we
present the implementation in [11], which modies the procedure for C-SVC.
7.1 The Decomposition Method
The decomposition method is an iterative process. In each step, the index set of variables
is partitioned to two sets B and N, where B is the working set. Then, in that iteration
variables corresponding to N are xed while a sub-problem on variables corresponding
to B is minimized. The procedure is as follows:
Algorithm 1 (Decomposition method)
1. Given a number q l as the size of the working set. Find
1
as an initial feasible
solution of (76). Set k = 1.
2. If
k
is an optimal solution of (76), stop. Otherwise, nd a working set B
1, . . . , l whose size is q. Dene N 1, . . . , lB and
k
B
and
k
N
to be sub-
vectors of
k
corresponding to B and N, respectively.
3. Solve the following sub-problem with the variable
B
:
minimize

B
R
q
1
2

iB,jB

j
y
i
y
j
k(x
i
, x
j
) +

iB,jN

k
j
y
i
y
j
k(x
i
, x
j
)
subject to 0
i

1
m
, i B, (78)

iB

i
y
i
=

iN

k
i
y
i
, (79)

iB

i
=

iN

k
i
. (80)
4. Set
k+1
B
to be the optimal solution of (78) and
k+1
N

k
N
. Set k k + 1 and
goto Step 2.
Note that B is updated in each iteration. To simplify the notation, we simply use B
instead of B
k
.
7.2 Working Set Selection
An important issue of the decomposition method is the selection of the working set B.
Here, we consider an approach based on the violation of the KKT condition. Similar to
(30), by putting (61) into (57), one of the KKT conditions is

i
[y
i
(
m

j=1

j
K(x
i
, x
j
) + b) +
i
] = 0, i = 1, . . . , m. (81)
Using 0
i

1
m
, (81) can be rewritten as:
m

j=1

j
y
i
y
j
k(x
i
, x
j
) + by
i
0, if
i
<
1
m
,
m

j=1

j
y
i
y
j
k(x
i
, x
j
) + by
i
0, if
i
> 0.
(82)
That is, an is optimal for the dual problem (76) if and only if is feasible and satises
(81). Using the property that y
i
= 1 and representing W()
i
=

m
j=1

j
y
i
y
j
K(x
i
, x
j
),
(82) can be further written as
max
iI
1
up
()
W()
i
b min
iI
1
low
()
W()
i
and
max
iI
1
up
()
W()
i
+ b min
iI
1
low
()
W()
i
,
(83)
where
I
1
up
() := i [
i
> 0, y
i
= 1, I
1
low
() := i [
i
<
1
m
, y
i
= 1, (84)
and
I
1
up
() := i [
i
<
1
m
, y
i
= 1, I
1
low
() := i [
i
> 0, y
i
= 1. (85)
We call any (i, j) I
1
up
() I
1
low
() or I
1
up
() I
1
low
() satisfying
y
i
W()
i
> y
j
W()
j
(86)
a violating pair as (83) is not satised. When is not optimal yet, if any such a violating
pair is included in B, the optimal objective value of (78) is small than that at
k
.
Therefore, the decomposition procedure has its objective value strictly decreasing from
one iteration to the next.
Therefore, a natural choice of B is to select all pairs which violate (83) the most.
To be more precise, we can set q to be an even integer and sequentially select q/2 pairs
(i
1
, j
1
), . . . , (i
q/2
, j
q/2
) from I
1
up
() I
1
low
() or I
1
up
() I
1
low
() such that
y
i
1
W()
i
1
y
j
1
W()
j
1
y
i
q/2
W()
i
q/2
y
j
q/2
W()
j
q/2
. (87)
This working set selection is merely an extension of that for C-SVC. The main
difference is that for C-SVM, (83) becomes only one inequality with b. Due to this
similarity, we believe that the convergence analysis of C-SVC [21] can be adapted here
though detailed proofs have not been written and published.
[11] considers the same working set selection. However, following the derivation
for C-SVC in [19], it is obtained using the concept of feasible directions in constrained
optimization. We feel that a derivation from the violation of the KKT condition is more
intuitive.
7.3 SMO-type Implementation
The Sequential Minimal Optimization (SMO) algorithm [28] is an extreme of the de-
composition method where, for C-SVC, the working set is restricted to only two el-
ements. The main advantage is that each two-variable sub-problem can be analyti-
cally solved, so numerical optimization software are not needed. For this method, at
least two elements are required for the working set. Otherwise, the equality constraint

iB

i
y
i
=

jN

k
j
y
j
leads to a xed optimal objective value of the sub-problem.
Then, the decomposition procedure stays at the same point.
Now the dual of -SVC possesses two inequalities, so we may think that more
elements are needed for the working set. Indeed, two elements are still enough for the
case of -SVC. Note that (79) and (80) can be rewritten as

iB,y
i
=1

i
y
i
=

2

iN,y
i
=1

k
i
y
i
and

iB,y
i
=1

i
y
i
=

2

iN,y
i
=1

k
i
y
i
. (88)
Thus, if (i
1
, j
1
) are selected as the working set selection using (87), y
i
1
= y
j
1
, so
(88) reduces to only one equality with two variables. Then, the sub-problem is still
guaranteed to be smaller than that at
k
.
The comparison in [11] shows that using C and with the connection in proposition
3 and equivalent stopping condition, the performance of the SMO-type implementation
described here for C-SVM and -SVM are comparable.
7.4 The Calculation of b and and Stopping Criteria
If at an optimal solution, 0 <
i
< 1/m and y
i
= 1, then i I
1
up
() and I
1
low
().
Thus, b = W()
i
. Similarly, if there is another 0 <
j
< 1/m and y
j
= 1,
then + b = W()
j
. Thus, solving two equalities gives b and . In practice, we
average W()
i
to avoid numerical errors:
b =

0<
i
<
1
m
,y
i
=1
W()
i

0<
i
<
1
m
,y
i
=1
1
, (89)
If there are no components such that 0 <
i
< 1/m, b (and + b) can be any
number in the interval formed by (83). A common way is to select the middle point and
then still solves two linear equations
The stopping condition of the decomposition method can easily followthe newform
of the optimality condition (83):
max
_
min
iI
1
low
()
W()
i
+ max
iI
1
up
()
W()
i
,
min
iI
1
low
()
W()
i
+ max
iI
1
up
()
W()
i
_
< ,
(90)
where > 0 is a chosen stopping tolerance.
8 Multi-Class -SV Classiers
Though SVM was originally designed for two-class problems, several approaches have
been developed to extend SVM for multi-class data sets. In this section, we discuss the
extension of the one-against-one approach for multi-class -SVM.
Most approaches for multi-class SVM decompose the data set to several binary
problems. For example, the one-against-one approach trains a binary SVM for any
two classes of data and obtains a decision function. Thus, for a k-class problem, there
are k(k1)/2 decision functions. In the prediction stage, a voting strategy is used where
the testing point is designated to be in a class with the maximum number of votes. In
[18], it was experimentally shown that for general problems, using C-SV classier,
various multi-class approaches give similar accuracy. However, the one-against-one
method is more efcient for training. Here, we will focus on extending it for -SVM.
Multi-class methods must be considered together with parameter-selection strate-
gies. That is, we search for appropriate C and kernel parameters for constructing a
better model. In the following, we restrict the discussion on only the Gaussian (radius
basis function) kernel k(x
i
, x
j
) = e
x
i
x
j

2
, so the kernel parameter is . With
the parameter selection considered, there are two ways to implement the one-against-
one method: First, for any two classes of data, the parameter selection is conducted
to have the best (C, ). Thus, for the best model selected, each decision function has
its own (C, ). For experiments here, the parameter selection of each binary SVM is
by a ve-fold cross-validation. The second way is that for each (C, ), an evaluation
criterion (e.g. cross-validation) combining with the one-against-one method is used
for estimating the performance of the model. A sequence of pre-selected (C, ) is tried
to select the best model. Therefore, for each model, k(k1)/2 decision functions share
the same C and .
It is not very clear which one of the two implementations is better. On one hand, a
single parameter set may not be uniformly good for all k(k 1)/2 decision functions.
On the other hand, as the overall accuracy is the nal consideration, one parameter set
for one decision function may lead to over-tting. [14] is the rst to compare the two
approaches using C-SVM, where the preliminary results show that both give similar
accuracy.
For -SVM, each binary SVM using data from the ith and the jth classes has an
admissible interval [
ij
min
,
ij
max
], where
ij
max
= 2 min(m
i
, m
j
)/(m
i
+ m
j
) according
to proposition 3. Here m
i
and m
j
are the number of data points in the ith and jth
classes, respectively. Thus, if all k(k 1)/2 decision functions share the same , the
admissible interval is
[max
i=j

ij
min
, min
i=j

ij
max
]. (91)
This set is non-empty if the kernel matrix is positive denite. The reason is that proposi-
tion 3 implies
ij
min
= 0, i ,= j, so min
i=j

ij
max
= 0. Therefore, unlike C of C-SVM,
which has a large valid range [0, ), for -SVM, we worry that the admissible interval
may be too small. For example, if the data set is highly unbalanced, min
i=j

ij
min
is very
small.
We redo the same comparison as that in [14] for -SVM. Results are in Table 2. We
consider multi-class problems tested in [18], where most of them are from the statlog
collection [25]. Except data sets dna, shuttle, letter, satimage, and usps, where test sets
are available, we separate each problem to 80% training and 20% testing. Then, cross
validation are conducted only on the training data. All other settings such as data scaling
are the same as those in [18]. Experiments are conducted using LIBSVM [10], which
solves both C-SVM and -SVM.
Results in Table 2 show no signicant difference among the four implementations.
Note that some problems (e.g. shuttle) are highly unbalanced so the admissible interval
(91) is very small. Surprisingly, from such intervals, we can still nd a suitable which
leads to a good model. This preliminary experiment indicates that in general the use of
one-against-one approach for multi-class -SVM is viable.
Table 2. Test accuracy (in percentage) of multi-class data sets by C-SVM and -SVM. The
columns Common C, Different C, Common , Different are testing accuracy of using
the same and different (C,), (or (,)) for all k(k 1)/2 decision functions. The validation is
conducted on the following points of (C,): [2
5
, 2
3
, . . . , 2
15
] [2
15
, 2
13
, . . . , 2
3
]. For -
SVM, the range of is the same but we validate a 10-point discretization of in the interval (91)
or [
ij
min
,
ij
max
], depending on whether k(k 1)/2 decision functions share the same parameters
or not. For small problems (number of training data 1000), we do cross validation ve times,
and then average the testing accuracy.
Data set Class No. # training # testing Common C Different C Common Different
vehicle 4 677 169 86.5 87.1 85.9 87.8
glass 6 171 43 72.2 70.7 73.0 69.3
iris 3 120 30 96.0 93.3 94.0 94.6
dna 3 2000 1186 95.6 95.1 95.0 94.8
segment 7 1848 462 98.3 97.2 96.7 97.6
shuttle 7 43500 14500 99.9 99.9 99.7 99.8
letter 26 15000 5000 97.9 97.7 97.9 96.8
vowel 11 423 105 98.1 97.7 98.3 96.0
satimage 6 4435 2000 91.9 92.2 92.1 91.9
wine 3 143 35 97.1 97.1 97.1 96.6
usps 10 7291 2007 95.3 95.2 95.3 94.8
We also present the contours of C-SVM and -SVM in Figure 8 using the approach
that all decision functions share the same (C, ). In the contour of C-SVM, the x-
axis and y-axis are log
2
C and log
2
, respectively. For -SVM, the x-axis is in the
interval (91). Clearly, the good region of using -SVM is smaller. This conrms our
concern earlier, which motivated us to conduct experiments in this section. Fortunately,
points in this smaller good region still lead to models that are competitive with those by
C-SVM.
There are some ways to enlarge the admissible interval of . A work to extend
algorithm to the case of very small values of by allowing negative margins is [27].
For the upper bound, according to the above proposition 3, if the classes are balanced,
then the upper bound is 1. This leads to the idea to modify the algorithm by adjusting
the cost function such that the classes are balanced in terms of the cost, even if they are
not in terms of the mere numbers of training examples. An earlier discussion on such
formulations is at [12]. For example, we can consider the following formulation:
minimize
wH,R
m
,,bR
(w, , ) =
1
2
|w|
2
+
1
2m
+

i:y
i
=1

i
+
1
2m

i:y
i
=1

i
subject to y
i
(x
i
, w) + b)
i
,
and
i
0, 0.
satimage 91.5
91
90.5
90
89.5
89
88.5
-5 0 5 10 15
log(C)
-16
-14
-12
-10
-8
-6
-4
-2
0
2
4
log(gamma)
satimage 91.5
91
90.5
90
89.5
89
88.5
0 0.1 0.2 0.3 0.4 0.5 0.6
NU
-16
-14
-12
-10
-8
-6
-4
-2
0
2
4
log(gamma)
Fig. 6. 5-fold cross-validation accuracy of the data set satimage. Left: C-SVM, Right: -SVM
The dual is
maximize
R
m
W() =
1
2
m

i,j=1

j
y
i
y
j
k(x
i
, x
j
)
subject to 0
i

1
2m
+
, if y
i
= 1,
0
i

1
2m

, if y
i
= 1,
m

i=1

i
y
i
= 0,
m

i=1

i
.
Clearly, when all
i
equals its corresponding upper bound, is a feasible solution with

m
i=1

i
= 1 .
Another possibility is
minimize
wH,R
m
,,bR
(w, , ) =
1
2
|w|
2
+
1
2 min(m
+
, m

)
m

i=1

i
subject to y
i
(x
i
, w) + b)
i
,
and
i
0, 0.
The dual is
maximize
R
m
W() =
1
2
m

i,j=1

j
y
i
y
j
k(x
i
, x
j
)
subject to 0
i

1
2 min(m
+
, m

)
,
m

i=1

i
y
i
= 0,
m

i=1

i
.
Then, the largest admissible is 1.
A slight modication of the implementation in Section 7 for the above formulations
is in [13].
9 Applications of -SV Classiers
Researchers have applied -SVM on different applications. Some of them feel that it
is easier and more intuitive to deal with [0, 1] than C [0, ). Here, we briey
summarize some work which use LIBSVM to solve -SVM.
In [7], researchers from HP Labs discuss the topics of personal email agent. Data
classication is an important component for which the authors use -SVM because they
think the parameter is more intuitive than the C parameter.
[23] applies machine learning methods to detect and localize boundaries of natural
images. Several classiers are tested where, for SVM, the authors considered -SVM.
10 Conclusion
One of the most appealing features of kernel algorithms is the solid foundation pro-
vided by both statistical learning theory and functional analysis. Kernel methods let
us interpret (and design) learning algorithms geometrically in feature spaces nonlin-
early related to the input space, and combine statistics and geometry in a promising
way. Kernels provide an elegant framework for studying three fundamental issues of
machine learning:
Similarity measures the kernel can be viewed as a (nonlinear) similarity measure,
and should ideally incorporate prior knowledge about the problem at hand
Data representation as described above, kernels induce representations of the
data in a linear space
Function class due to the representer theorem, the kernel implicitly also deter-
mines the function class which is used for learning.
Support vector machines have been one of the major kernel methods for data classi-
cation. Its original form requires a parameter C [0, ), which controls the trade-off
between the classier capacity and the training errors. Using the -parameterization,
the parameter C is replaced by a parameter [0, 1]. In this tutorial, we have given its
derivation and present possible advantages of using the -support vector classier.
Acknowledgments
The authors thank Ingo Steinwart and Arthur Gretton for some helpful comments.
References
1. M. A. Aizerman,

E.. M. Braverman, and L. I. Rozono er. Theoretical foundations of the
potential function method in pattern recognition learning. Automation and Remote Control,
25:821837, 1964.
2. N. Alon, S. Ben-David, N. Cesa-Bianchi, and D. Haussler. Scale-sensitive dimensions, uni-
form convergence, and learnability. Journal of the ACM, 44(4):615631, 1997.
3. M. Avriel. Nonlinear Programming. Prentice-Hall Inc., New Jersey, 1976.
4. P. L. Bartlett and J. Shawe-Taylor. Generalization performance of support vector machines
and other pattern classiers. In B. Sch olkopf, C. J. C. Burges, and A. J. Smola, editors,
Advances in Kernel Methods Support Vector Learning, pages 4354, Cambridge, MA,
1999. MIT Press.
5. M. S. Bazaraa, H. D. Sherali, and C. M. Shetty. Nonlinear programming : theory and algo-
rithms. Wiley, second edition, 1993.
6. K. P. Bennett and E. J. Bredensteiner. Duality and geometry in SVMclassiers. In P. Langley,
editor, Proceedings of the 17th International Conference on Machine Learning, pages 5764,
San Francisco, California, 2000. Morgan Kaufmann.
7. R. Bergman, M. Griss, and C. Staelin. A personal email assistant. Technical Report HPL-
2002-236, HP Laboratories, Palo Alto, CA, 2002.
8. B. E. Boser, I. M. Guyon, and V. Vapnik. A training algorithm for optimal margin classiers.
In D. Haussler, editor, Proceedings of the 5th Annual ACM Workshop on Computational
Learning Theory, pages 144152, Pittsburgh, PA, July 1992. ACM Press.
9. C. J. C. Burges and B. Sch olkopf. Improving the accuracy and speed of support vector
learning machines. In M. Mozer, M. Jordan, and T. Petsche, editors, Advances in Neural
Information Processing Systems 9, pages 375381, Cambridge, MA, 1997. MIT Press.
10. C.-C. Chang and C.-J. Lin. LIBSVM: a library for support vector machines, 2001. Software
available at http://www.csie.ntu.edu.tw/cjlin/libsvm.
11. C.-C. Chang and C.-J. Lin. Training -support vector classiers: Theory and algorithms.
Neural Computation, 13(9):21192147, 2001.
12. H.-G. Chew, R. E. Bogner, and C.-C. Lim. Dual -support vector machine with error rate
and training size biasing. In Proceedings of ICASSP, pages 126972, 2001.
13. H. G. Chew, C. C. Lim, and R. E. Bogner. An implementation of training dual-nu support
vector machines. In Qi, Teo, and Yang, editors, Optimization and Control with Applications.
Kluwer, 2003.
14. K.-M. Chung, W.-C. Kao, C.-L. Sun, and C.-J. Lin. Decomposition methods for linear sup-
port vector machines. Technical report, Department of Computer Science and Information
Engineering, National Taiwan University, 2002.
15. C. Cortes and V. Vapnik. Support vector networks. Machine Learning, 20:273297, 1995.
16. D. J. Crisp and C. J. C. Burges. A geometric interpretation of -SVM classiers. In S. A.
Solla, T. K. Leen, and K.-R. M uller, editors, Advances in Neural Information Processing
Systems 12. MIT Press, 2000.
17. A. Gretton, R. Herbrich, O. Chapelle, B. Sch olkopf, and P. J. W. Rayner. Estimating the
Leave-One-Out Error for Classication Learning with SVMs. Technical Report CUED/F-
INFENG/TR.424, Cambridge University Engineering Department, 2001.
18. C.-W. Hsu and C.-J. Lin. A comparison of methods for multi-class support vector machines.
IEEE Transactions on Neural Networks, 13(2):415425, 2002.
19. T. Joachims. Making largescale SVM learning practical. In B. Sch olkopf, C. J. C. Burges,
and A. J. Smola, editors, Advances in Kernel Methods Support Vector Learning, pages
169184, Cambridge, MA, 1999. MIT Press.
20. C.-J. Lin. Formulations of support vector machines: a note from an optimization point of
view. Neural Computation, 13(2):307317, 2001.
21. C.-J. Lin. On the convergence of the decomposition method for support vector machines.
IEEE Transactions on Neural Networks, 12(6):12881298, 2001.
22. Luntz, A. and Brailovsky, V. On estimation of characters obtained in statistical procedure of
recognition. Technicheskaya Kibernetica 3, 1969.
23. D. R. Martin, C. C. Fowlkes, and J. Malik. Learning to detect natural image boundaries using
brightness and texture. In Advances in Neural Information Processing Systems, volume 14,
2002.
24. J. Mercer. Functions of positive and negative type and their connection with the theory of
integral equations. Philosophical Transactions of the Royal Society, London, A 209:415
446, 1909.
25. D. Michie, D. J. Spiegelhalter, and C. C. Taylor. Machine Learning, Neural and Sta-
tistical Classication. Prentice Hall, Englewood Cliffs, N.J., 1994. Data available at
http://www.ncc.up.pt/liacc/ML/statlog/datasets.html.
26. M. Opper and O. Winther. Gaussian processes and SVM: Mean eld and leave-one-out
estimator. In A. Smola, P. Bartlett, B. Sch olkopf, and D. Schuurmans, editors, Advances in
Large Margin Classiers, Cambridge, MA, 2000. MIT Press.
27. F. Perez Cruz, J. Weston, D. J. L. Herrmann, and B. Sch olkopf. Extension of the -svm
range for classication. In J. Suykens, G. Horvath, S. Basu, C. Micchelli, and J. Vandewalle,
editors, Advances in Learning Theory: Methods, Models and Applications, 190, pages 179
196, Amsterdam, 2003. IOS Press.
28. J. C. Platt. Fast training of support vector machines using sequential minimal optimization.
In B. Sch olkopf, C. J. C. Burges, and A. J. Smola, editors, Advances in Kernel Methods -
Support Vector Learning, Cambridge, MA, 1998. MIT Press.
29. B. Sch olkopf. Support Vector Learning. R. Oldenbourg Verlag, M unchen, 1997. Doktorar-
beit, Technische Universit at Berlin. Available from http://www.kyb.tuebingen.mpg.de/bs.
30. B. Sch olkopf, C. J. C. Burges, and A. J. Smola. Advances in Kernel Methods Support
Vector Learning. MIT Press, Cambridge, MA, 1999.
31. B. Sch olkopf and A. J. Smola. Learning with Kernels. MIT Press, Cambridge, MA, 2002.
32. B. Sch olkopf, A. J. Smola, R. C. Williamson, and P. L. Bartlett. New support vector algo-
rithms. Neural Computation, 12:12071245, 2000.
33. A. J. Smola, P. L. Bartlett, B. Sch olkopf, and D. Schuurmans. Advances in Large Margin
Classiers. MIT Press, Cambridge, MA, 2000.
34. I. Steinwart. Support vector machines are universally consistent. Journal of Complexity,
18:768791, 2002.
35. I. Steinwart. On the optimal parameter choice for -support vector machines. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, 2003. To appear.
36. I. Steinwart. Sparseness of support vector machines. Technical report, 2003.
37. V. Vapnik. Estimation of Dependences Based on Empirical Data (in Russian). Nauka,
Moscow, 1979. (English translation: Springer Verlag, New York, 1982).
38. V. Vapnik. The Nature of Statistical Learning Theory. Springer Verlag, New York, 1995.
39. V. Vapnik. Statistical Learning Theory. Wiley, NY, 1998.
40. V. Vapnik and O. Chapelle. Bounds on error expectation for support vector machines. Neural
Computation, 12(9):20132036, 2000.
41. V. Vapnik and A. Chervonenkis. Theory of Pattern Recognition (in Russian). Nauka,
Moscow, 1974. (German Translation: W. Wapnik & A. Tscherwonenkis, Theorie der Ze-
ichenerkennung, AkademieVerlag, Berlin, 1979).
42. V. Vapnik and A. Lerner. Pattern recognition using generalized portrait method. Automation
and Remote Control, 24:774780, 1963.
43. G. Wahba. Spline Models for Observational Data, volume 59 of CBMS-NSF Regional Con-
ference Series in Applied Mathematics. Society for Industrial and Applied Mathematics,
Philadelphia, Pennsylvania, 1990.
44. R. C. Williamson, A. J. Smola, and B. Sch olkopf. Generalization performance of regulariza-
tion networks and support vector machines via entropy numbers of compact operators. IEEE
Transactions on Information Theory, 47(6):25162532, 2001.
45. P. Wolfe. A duality theorem for non-linear programming. Quartely of Applied Mathematics,
19:239244, 1961.