TELKOM
NIKA
, Vol.14, No
.2, June 20
16
, pp. 638~6
4
6
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v14i1.2751
638
Re
cei
v
ed
De
cem
ber 2
0
, 2015; Re
vi
sed
March 18, 20
15; Accepted
April 8, 2016
Semi-supervised Online Multiple Kernel Learning
Algorithm for Bi
g Data
Ning Liu
1
*, Jianhua Zh
ao
2
1
School of ec
o
nomics a
nd ma
nag
ement, Sha
ngl
uo Un
iversit
y
, Sha
ngl
uo 7
2
600
0, Chi
n
a
2
School of mat
hematics a
nd c
o
mputer a
p
p
lic
ations
, Sha
n
g
l
uo Un
iversit
y
, Shan
glu
o
72
60
00, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: liuni
ng
20
122
014
@al
i
y
un.c
o
m
A
b
st
r
a
ct
In order to i
m
prove the
perf
o
rmanc
e of machi
ne
le
arn
i
n
g
in bi
g dat
a, onli
ne
mu
ltipl
e
kerne
l
lear
nin
g
al
gorit
hms
are pr
opo
sed in th
is pa
p
e
r. F
i
rst,
a supervise
d onl
in
e mu
ltipl
e
kern
el
lear
nin
g
al
gorit
h
m
for big data (S
OMK_bd) is pr
opos
ed to red
u
ce the co
m
p
u
t
ationa
l w
o
rklo
ad duri
ng ker
n
el modific
a
tio
n
. In
SOMK_bd, the
tradition
al ker
nel l
earn
i
n
g
al
gorith
m
is
i
m
pr
oved a
nd ker
n
el inte
gratio
n is
only carri
ed o
u
t in
the co
nstructe
d kern
el s
ubs
e
t. Next, an u
n
s
uperv
i
sed
on
lin
e
multi
p
le
kern
el l
earn
i
n
g
a
l
g
o
rith
m for
big
d
a
ta
(UOMK_bd) is
prop
osed. In U
O
MK_bd, the traditi
ona
l kern
el
learni
ng a
l
gor
i
t
hm is i
m
pr
ove
d
to adapt to the
onli
ne
envir
on
me
nt an
d data
replac
e
m
e
n
t strategy is use
d
to mo
dify the
kerne
l
functio
n
in uns
up
ervis
e
d
ma
nn
er. T
hen,
a se
mi-su
per
vised
on
lin
e
multipl
e
k
e
rn
e
l
l
earn
i
ng
al
gor
ithm for b
i
g
dat
a (SSOMK_b
d
)
is
prop
osed. Bas
ed on i
n
cre
m
e
n
tal le
arni
ng, SSOMK_bd mak
e
s full use of the ab
und
ant
in
formati
on of lar
g
e
scale i
n
co
mp
le
te labe
le
d dat
a, and us
es S
O
MK_bd
a
nd
UOMK_bd to
upd
ate the cur
r
ent read
in
g d
a
ta.
F
i
nally,
exp
e
ri
me
nts ar
e c
o
n
ducted
o
n
U
C
I
data s
e
t a
n
d
th
e res
u
lts s
how
that the
pro
pos
ed
alg
o
rith
ms
are
effective.
Ke
y
w
ords
: Se
mi-s
uperv
i
se
d Classific
a
tio
n
, Onlin
e Le
arni
n
g
, Multipl
e
Ker
nel, Big D
a
ta
Copy
right
©
2016 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
In rece
nt years, with the ra
pid develop
m
ent of
the Internet, clo
ud computing, Internet o
f
things, soci
al
media a
nd ot
her info
rmatio
n techn
o
logy,
data from all
walks of life
are
sho
w
in
g an
explosive g
r
owth tre
nd [1]. We have
entere
d
the
era of big
data, whi
c
h
has b
e
come
an
importa
nt st
rategic re
so
urce
of the
co
untry,
the
m
anag
ement a
nd
a
nalysi
s
of
big data ha
s
become a hot
topic of aca
d
e
mic an
d ind
u
strial attenti
on [2, 3].
The purpo
se of
data coll
ection,
data sto
r
age,
data
tra
n
smi
ssi
on
an
d data
ma
na
gemen
t
is to use big
data effective
l
y, where m
a
chin
e lear
nin
g
techn
o
logy
is esse
ntial [4, 5]. In rece
nt
years, m
a
chi
ne lea
r
nin
g
in
big data
bein
g
one
of
the hotsp
ots
h
a
s attracte
d extensive attentio
n,
and ne
w a
c
hi
evements
are eme
r
ging [
7
, 8]. For exam
ple, Klein
e
r et al., [9] put forward a
ne
w
data
sam
p
lin
g meth
od
of BLB b
a
se
d
Baggi
ng
l
earni
ng th
ou
ght, to
solve
the
bottlen
eck
probl
em
s of cal
c
ulatio
n in big dat
a Bootstra
p; Go
nzal
ez et al.
,
[10] presen
ted distribut
ed
machi
ne l
e
a
r
ning f
r
ame
w
o
r
k graph
to
realize the
la
rge-scale
ma
chine l
earning;
Gao
et
al., [11]
prop
osed the
idea of "one
pass learnin
g
" lear
nin
g
, trying to only sca
n again
the data usin
g
con
s
tant
storage to sto
r
e i
n
terme
d
iate result
s,
whi
c
h
achi
eved go
o
d
re
sults in A
UC o
p
timization
su
ch a compl
e
x learnin
g
task.
However, there are
still many probl
ems to
be solved in machine l
earni
ng for bi
g data
due to it’s complexity [6, 7]. From the view of
the algorithm,
it mainly exists the follo
wing
probl
em
s in
machi
ne l
earning a
nd th
e
analysi
s
of
big data
mini
ng. Becau
s
e
of the hu
ge
data
size, it is not within the
accepta
b
le time to
get results. So putting forwa
r
d
a new ma
chi
n
e
learni
ng algo
rithm to meet
the deman
d of high
data pro
c
e
ssi
ng a
nd larg
e data
is one of the hot
r
e
sear
ch points
in mach
ine learning [3, 7].
On this i
s
sue,
the resea
r
ch
ers
gen
erally
solve the p
r
o
b
lem of big d
a
ta machine l
earni
ng
throug
h two
method
s. A method is to
modify ex
isting machi
ne le
arnin
g
algo
rithm and tra
n
sform
it to be concurrent\parallel
computin
g versi
on[6,10]; the other me
thod is to design a onlin
e
learni
ng versi
on of existing machi
ne le
arnin
g
and d
a
t
a mining alg
o
rithm [6, 7].
Becau
s
e th
e online le
arni
n
g
doe
sn't ne
e
d
to store
sa
mples
of earl
y
data, or onl
y needs
to save the early data from a sam
p
le
of a sufficie
n
t statistic, it is very suitable for big data
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Sem
i
-supe
rvi
s
ed O
n
line M
u
ltiple Kern
el Learning Alg
o
rithm
for Big Data (Ning Li
u)
639
analysi
s
ap
pli
c
ation
scena
rios. Fo
r very l
a
rge
dat
a
set
s
, online l
earning get
s dat
a in a sequ
en
tial
manne
r and
synchro
nou
s
update lea
r
ni
ng; for hi
gh speed dat
a streams, onli
ne
learni
ng can
be
carrie
d out while data is i
nputting. And
the lear
nin
g
model can reflect a re
ce
nt period
of time
input data
rul
e
and
ma
ke
effective pred
iction [12,
1
3
]. In rece
nt years, some
rese
arche
r
s p
u
t
online lea
r
ni
n
g
into the field of machin
e
lear
nin
g
and
produ
ce
d be
tter benefits. For exampl
e,
Shalev-Sh
w
a
r
tz S [1
4] p
r
o
posed
gra
d
ie
nt asce
nt
(d
e
s
cent)
metho
d
ba
se
d o
n
t
he id
ea
of on
line
learni
ng to impleme
n
t a large
-
scale m
odel of fa
st learni
ng; Yan
g
[15] prop
osed an in
cre
m
ental
optimizatio
n fast de
cisi
on tree alg
o
rithm
for data with
noise. Comp
ared
with the
traditional bi
g
data mini
ng
deci
s
io
n tre
e
algo
rithm, th
e main
adv
a
n
tage of thi
s
algo
rithm
was
with real
-time
mining capa
b
ilities, which coul
d
store the com
p
lete
data for tr
aini
ng de
cisio
n
model when
the
mobile data
stream was inf
i
nite.
As a kin
d
of importa
nt onli
ne lea
r
ning
method,
onli
n
e multiple kernel lea
r
ning
h
a
s be
en
widely used
in different application fields [16-19
]. Ho
wever, du
e to the indirect influe
nce of
kernel l
earni
ng p
r
obl
em
on target an
alysis, o
n
line
multiple ke
rnel
lea
r
nin
g
has not
b
een
fully
resea
r
ched in
the field of big data analy
s
is and mini
ng
[7].
In orde
r to improve the p
e
rforman
c
e of
machi
ne lea
r
ning in big d
a
t
a environm
e
n
t, semi-
sup
e
rvised l
e
arnin
g
a
nd
o
n
line m
u
ltiple
ke
rnel
lea
r
ni
ng a
r
e i
n
tro
d
u
ce
d into
the
field of
big
d
a
ta
machi
ne lea
r
ning in this p
aper. Fi
rst
,
a
supe
rvise
d
o
n
line multiple
kernel lea
r
ni
ng algo
rithm
for
big d
a
ta
(SO
M
K_bd) i
s
p
r
opo
sed
to
re
duce th
e
co
mputational
worklo
ad du
ri
ng
o
n
line
lea
r
ning
kernel mo
dification. Next,
an un
sup
e
rvi
s
ed o
n
line
m
u
ltiple ke
rnel
learni
ng alg
o
rithm (UO
M
K_
bd)
is p
r
op
osed
to adapt
to
the onli
ne l
earni
ng
environment. T
h
e
n
, an
semi
-supervi
sed
on
line
multiple
ke
rn
el lea
r
nin
g
al
gorithm
for
bi
g dat
a
(SSO
MK_bd) is p
r
opo
sed. Ba
sed o
n
the
onl
ine
learni
ng fra
m
ewo
r
k of increme
n
tal lear
ni
ng, SSOMK_bd ma
ke
s full use
of the abundant
informatio
n o
f
large scale
labele
d
and
unlab
el
ed dat
a, and uses t
he SOMK_bd
algorithm an
d
UOMK_
bd
al
gorithm
to u
pdate th
e
cu
rre
nt re
adin
g
data
se
para
t
ely. Finally, experim
ents
are
con
d
u
c
ted on
the benchma
r
k
UCI data
set to verify
th
e effectivene
ss of the p
r
op
ose
d
algo
rith
m.
The
stru
cture
of this pap
e
r
is a
s
follo
ws:
th
e
part
I is th
e introdu
ction; the
part II
introduces the propo
sed algorithms: SOMK_bd, UOMK_bd
and SSOMK_bd; the part III is the
experim
ent a
nd analy
s
is; the part IV is the co
ncl
u
si
on
; the part V is the ackno
w
le
dgeme
n
ts.
2. Proposed
Algorithms
In this p
ape
r, we
pro
p
o
s
e
three
kin
d
s
o
f
algorithm
s i
n
big
data e
n
v
ironme
n
t, there
are
sup
e
rvised o
n
line multiple
kernel lea
r
n
i
ng al
go
rithm
for big data
(SOMK_bd
),
unsu
pervi
se
d
online m
u
ltipl
e
ke
rnel l
earning alg
o
rith
m for big d
a
ta (UOMK_b
d
)
and
semi
-supervi
sed
onl
ine
multiple ke
rn
el learni
ng al
gorithm for bi
g data
(SSO
MK_bd). Th
e three alg
o
rith
ms are de
scri
bed
in the followin
g
part of the pape
r.
2.1. Supervi
sed Online
Multiple Kernel L
earning
Algorithm for Big Da
ta (SOMK_
bd)
The m
a
in p
u
rpose of m
u
ltiple kernel le
a
r
ning
is t
o
stu
d
y ke
rnel
fun
c
tion
with p
a
rametri
c
or se
mi-p
ara
m
etric di
re
ctly from
the training data,
whi
c
h contai
n the informa
t
ion reflectin
g
the
data di
strib
u
tion. We ta
ke i
t
as
given
a
set of traini
ng
data
D
L
= { (
x
i
,
y
i
)|i = 1
,,
,
,
2
…
n} wh
ere
x
i
is feature
set, y
i
∈-
,
,
{
1
+1} i
s
th
e cl
a
s
s label
K
m
= {
k
j
(·
,R
,
,
,
,
·)
: X ×
X
→
j
=
1
2
…
m
}
i
s
given a set contai
ning m
basi
c
kernel
function
s
,
u =
{
u
1
,
u
2
,
…u
m
,
∑
u
i
= 1
}
is a
set of no
n
negative wei
ghts
,
mini
mizing the
cla
ssi
fication e
r
ror
of t
he ke
rnel
learni
ng ma
chine o
n
the t
e
st
set. Beca
use
of non ne
gat
ive weight
s, convex com
b
i
nation
kernel
function i
s
still a valid ke
rn
el
function. The
probl
em can
be formali
z
e
d
as the formul
a(1
)
:
2
ii
1
mi
n
(f(x
)
,
y
)
k
k
n
fH
H
i
fC
l
(1)
For the formu
l
a (1), it is difficult to dire
ctl
y
compute th
e optimal val
ue ,even it is basi
c
ally
impossibl
e
wi
thin a
c
cepta
b
le time
to fi
nd the
o
p
timal solution,
e
s
pe
cially
und
er th
e bi
g
da
ta
environ
ment .
In general, online multiple
kern
el learni
ng
will tran
sf
orm the opti
m
ization p
r
o
b
lem of
formula
(1) in
to the followi
ng problem: first, find out the optimal fu
nction f
i
of e
a
ch
ke
rnel K
i
in
their resp
ecti
ve Hilbe
r
t sp
ace
H
ki
; then,
look fo
r a set of weight
s u
i
,
which ma
ke
s the f
i
to be the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 638 – 64
6
640
best
co
mbin
a
t
ion, and
u
p
d
a
te weight
s
u
i
and
f
i
synchronou
sly in
th
e p
r
o
c
e
s
s of
sea
r
ching
for
the
optimal value
.
Whe
n
the co
mbination of kernel
fu
nctio
n
is
lin
ea
r, the onli
ne m
u
ltiple ke
rnel
l
earni
ng
probl
em can
be solved by the
followi
ng three step
s: i
n
the first
ste
p
, train
a
classifier f
i
usi
ng
t
he
basi
s
of the t
r
aining
set for
each kern
el functio
n
.
The
se
con
d
ste
p
i
s
to pe
rform
online l
earnin
g
.
After readi
ng
a trainin
g
sa
mple, differe
n
t
strategi
e
s
a
r
e ado
pted to
modify the ke
rnel
weig
ht a
n
d
the cl
assifier
according
to
different p
r
e
d
i
ction
re
sults.
In the thi
r
d
step, ite
r
ative
until it m
eet
s
certai
n
con
d
ition, the o
p
timal kern
el fu
nction i
s
a weighted
co
mb
ination of
ea
ch ke
rn
el fun
c
tion
with the o
p
timal wei
ght. Duri
ng the
seco
nd
step
and the thi
r
d step,
kern
el wei
ghts a
n
d
cla
ssifie
r
s’
up
dating
strate
g
y
is a
s
follo
ws: firs
t
rea
d
a
training
sam
p
le, then d
e
termin
e wheth
e
r
the predictio
n of the
sam
p
le is
co
rrect
.
If corr
ect, d
o
not pe
rform any up
dati
ng a
c
tion; if
not,
update the
co
rre
sp
ondi
ng kernel
weig
ht and cl
assifier.
Ho
wever, the
r
e is a p
a
rticu
l
arly larg
e nu
mber
of kern
el function
s in
the pro
c
e
ss
of online
multiple
ke
rn
el lea
r
nin
g
.
Whe
n
e
a
ch
sample i
s
i
npu
t, all ke
rnel
fu
nction
are u
s
ed to
pre
d
ict
and
weig
ht; once
the fore
ca
st
is not
co
nsi
s
tent with
t
h
e co
rrect la
b
e
l, the wei
g
ht of all kernel
function
s a
r
e
neede
d to cha
nge a
nd
modify. This
will be a
wa
ste of co
mp
uting re
so
urces,
esp
e
ci
ally in big data environment,
it
s ef
f
i
cien
cy
is v
e
r
y
low.
In orde
r to improve the op
erating effici
e
n
cy
of online
multiple ke
rn
el learni
ng un
der th
e
environ
ment
of big d
a
ta a
nd redu
ce it
s com
put
ation
a
l re
so
urce
d
u
ring
modifyi
ng weight
an
d
cla
ssifie
r
, we improve the traditional onli
n
e multip
le ke
rnel learning a
l
gorithm , and
put forward a
sup
e
rvised o
n
line multiple
kernel lea
r
nin
g
algor
ith
m
for big data whi
c
h is n
a
med
as SOMK_
b
d
.
In the integ
r
a
t
ion process
of online
mul
t
iple ke
rnel
le
arnin
g
, SOM
K
_bd u
s
e
s
B
e
rno
u
lli
sampli
ng
to do
a random sampling. Only
the
samp
l
e
whose
sel
e
cti
on probability is 1 is sel
e
ct
ed
to be con
s
tru
c
ted a
s
a
su
bset of
kerne
l
function
and
only the one
in the sub
s
e
t
is to ca
rry o
u
t
function p
r
edi
ction, wei
ghte
d
combi
natio
n, kern
el
wei
ght and cl
assi
fier updatin
g, redu
cin
g
ke
rn
el
cal
c
ulatio
n workl
oad.
Algorithm 1. The Algorith
m
Description
of SOMK_bd
Inpu
t
:
– kernel functi
on set
:
K
m
= {
k
1
,
k
2
,,
…k
m
}
–
T
he
t
-th labe
led sam
p
le
:(
x
t
,y
t
)
– Initializ
ed cl
assifier
:
F = {
f
1
(t)
,
f
2
(t
)
,,
…f
m
(t)}
– W
e
ight
:
u
i
(t)=1 , i=1,.
..,m
– Discount factor
:∈
β
(0,1)
– Smoothin
g
p
a
rameter
:∈
δ
(0
,1
)
Outp
ut
:
kernel w
e
ig
ht
u
i
(t+
1
) and cl
assifier
f
i
(t+1),
i=1,.
..,m
Proc
e
dure
:
1
)
q
i(t)
=u
i
(t)/[max
1
≤
j
≤
m
u
j
(t)], i=1,
...
,
m
/*Kernel
w
e
i
ght prob
ab
ilit
y
q
i
(t)*/
2
)
p
i
(t)=(1
−δ
)q
i
(t)+
δ
/m, i=1,.
..,m
/*Kernel se
le
ction pro
b
a
b
il
ity p
i
(t)*/
3
)
for i=1,2,..
.,
m do
4
)
Sample s
i
(t)=
Bernoul
li_S
a
mpli
ng(
p
i
)
/*s
i
(t)=
1 represents i-th kerne
l
is selected, th
e one
w
h
er
e s
i
(t)=
1construct a
subset*/
5
)
end for
6
)
*
1
(
)
(
(
t
)
(t
)
s
i
g
n
(
(
t
)))
m
ii
i
i
yt
s
i
g
n
s
q
f
/*W
e
ightin
g in
the subset
a
n
d
predict the la
b
e
l of x
t
*/
7
)
for i=1,2,..
.,
m do
8
)
if
y
*
(t)=
y
t
then /*correct, not update
*
/
9
)
z
i
(t
)
=
0
10
)
else
/*not correct, upad
e*/
11
)
z
i
(t)=1
12
)
end if
13
)
Update u
i
(t+1)=u
i
(t)
β
z
i
(
t
)
m
i(
t)
/*
w
e
i
ght up
datin
g
w
h
ere z
i
(t)=
1 and m
i
(t)=1*/
14
)
Update f
i
(t+1)= f
i
(t)
+ z
i
(t
)
m
i
(t) y
t
k
i
(x
t
,•)
/*classifer u
pdati
ng
w
h
ere
z
i
(t)=
1 and m
i
(t)=1*/
15
)
en
d for
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Sem
i
-supe
rvi
s
ed O
n
line M
u
ltiple Kern
el Learning Alg
o
rithm
for Big Data (Ning Li
u)
641
The d
e
tailed
pro
c
ed
ure of
the alg
o
rith
m SOMK_bd
is
sho
w
n i
n
Algorithm
1. For th
e
input labele
d
sampl
e
s (x
t
,y
t
), in step 1) the
weighted probability
q
i
(t) is cal
c
ulate
d
; in step 2) the
i-th kernel sele
ction probability p
i
(t) is calcul
ated. He
re, p
i
(t) add
s the smo
o
thing
paramete
r
δ
to
ensure that each
kernel is
se
lected
with the probabilit
y of
δ
/m, and avoid that the proba
bility p
i
(t)
is con
c
entrated on
a few
kernel fun
c
tio
n
; By step 4) Berno
u
lli sa
mpling i
s
carried o
u
t and
th
e
probability s
i
(t) is
cal
c
ulate
d
. The o
ne
where
s
i
(t)
= 1
rep
r
e
s
ent
s t
hat the i-th
kernel
is
sel
e
cted
whe
n
the
t-th
sampl
e
i
s
i
n
p
u
t. Then, th
e
all kernel fu
n
c
tion
sel
e
cte
d
by Be
rnoulli
sampli
ng
(s
i
(t
)
=
1)
c
o
ns
titute
a s
u
bset; In the s
t
ep 6), the k
e
rnel
is
we
ighted in th
e
sub
s
et a
nd th
e predict
ed la
bel
of X is calcul
ated; Step 13) and
step
14) is th
e up
dating proce
dure of the
kernel
weig
ht and
kernel
cla
ssifi
er. He
re, the
updatin
g only
focu
s on t
h
e
sub
s
et of o
n
l
i
ne multiple
kernel l
earning
,
whe
r
e s
i
(t
)=1; That i
s
to
sa
y that we
onl
y update
th
e kernel wei
ght
and cla
s
sifie
r
where z
i
(t)
=
1
and m
i
(t)=1 in
SOMK_bd al
gorithm.
2.2. Unsuper
v
ised Online Multiple Kernel
Lear
ning Algorith
m
for Big Da
ta (UOM
K_b
d
)
Traditio
nal kernel le
arni
n
g
method b
a
se
d on da
ta depend
en
ce is a co
mmonly
unsupe
rvise
d
ke
rnel l
e
a
r
n
i
ng meth
od,
only co
nsi
d
e
r
ing th
e de
n
s
ity of data
distrib
u
tion. I
n
essen
c
e, it is to modify the ke
rnel fu
nction in
the tra
i
ning
set. It can modify an
y existing kern
e
l
function b
a
se
d on the ob
served d
a
ta sample
s , the essen
c
e i
s
to
modify the inner
pro
d
u
c
t on
the Hilbe
r
t sp
ace in
du
ced
by the kernel
function.
In gene
ral, the
kernel fun
c
tio
n
is modifie
d
by
the formula (2), whi
c
h ma
ke
s the dist
ribut
ion of the D in the data
set.
1
(
a
,
b
)
(
a,
b)
(
I
)
T
Da
D
b
Kk
k
M
KM
K
(2)
Whe
r
e D
=
{x
1
,x
2
,…,x
n
}; K is kernel fun
c
tion
,
k
xi
=(
k(x
i
,x
1
),k(x
i
,x
2
),…, k(x
i
,x
n
))
), K rep
r
e
s
ent
s th
e
Gram mat
r
ix of k about D
;
a and b are
2 training sa
mples; W
ij
=RBF(
x
i
,x
j
), Where x
i
and x
j
are
the element
s on the D , rep
r
esents
a sy
mmet
r
ic dist
a
n
ce
mat
r
ix
.
The cal
c
ul
ation of formula
(2) nee
ds to
be
in the off-line bat
ch, and the com
p
utational
time com
p
lex
i
ty is high. At the sa
me ti
me,
the kern
el functio
n
’s
updatin
g of a
and b
nee
d
to
comp
ute K
a
a
nd
K
b
in
the
data
sam
p
le
s. For large
d
a
ta st
ream
s,
M an
d K
D
i
s
in chan
ging, a
nd
not ea
sy to
calcul
ate. In
a
ddition, di
re
ct
cal
c
ul
ating t
he
whol
e d
a
ta set M
an
d
K
D
is
not real
istic
in comp
uting
resou
r
ces.
In order to m
a
ke the data
depe
nd on kernel le
arni
n
g
method suitable for onli
ne ke
rnel
modificatio
n
in big data
environ
ment.
We pr
ovide
an unsupe
rvised onli
n
e
multiple ke
rnel
learni
ng alg
o
rithm for big d
a
ta, which is
named a
s
UOMK_bd.
Algorithm 2. The Algorith
m
Description
of UOMK_bd
Input
:
–D=
{
x
1
,x
2
,…,x
n
}
–Curre
nt input
sample
:
x
t
–Gram matrix
and d
i
stanc
e matrix
:
K
,
M
Output
:
Updat
ed kern
el matri
x
:
K
proce
dure
:
1
)
Initializati
o
n
K
x0
2
)
K
xi
=(
k
(x
i
,x
1
),
k
(x
i
,x
2
),…,
k
(x
i
,x
n
)))
3
)
for j=
1,…,N
do
4
)
K
2
=
K
(j, •)
5
)
1
11
2
(I
)
t
T
x
Kk
k
M
KM
K
6
)
end for
7
)
Using
K
xt
to update matri
x
K in the last ro
w
a
nd the l
a
st colum
n
/*FI
FO displacement*/
8) return
K
The al
gorith
m
de
scription
of UOMK_b
d is
sh
own in
Algorithm
2.
UOMK_
bd fo
cu
se
s
o
n
the onli
ne u
p
dating
of M a
nd K
D
. In order to facilitate the
cal
c
ulati
on of M
and
K
D
, the si
ze
o
f
the
M and K
D
a
r
e
re
stricte
d
to
be fixed N*
N.
In orde
r to m
a
ke full
use o
f
information
of all of the d
a
ta
sampl
e
s to p
e
rform
ke
rnel
learnin
g
, the sample d
a
ta
repla
c
em
ent
strategy in D is de
sign
e
d
to
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 638 – 64
6
642
ensure that th
e other
sampl
e
s can be
co
me the el
eme
n
t of D. Beca
use the d
a
ta gene
ration
rul
e
s
of big
data
m
a
y ch
ang
e
wit
h
the
chan
ge
of time, we
l
e
arn
from the
operating sy
stem “first in first
out” (FIFO
)
p
age re
pla
c
e
m
ent policy t
o
repla
c
e th
e
data sam
p
le
s in D. He
re,
the timeliness of
data is con
s
id
ered.
In UOMK_b
d
,
it maintains
a wo
rking
set
of M whi
c
h
can be
used a
s
a
ca
che to
reflect
data di
strib
u
tion in
a pe
rio
d
time. The
wo
rkin
g
set lim
iti
ng st
rategy h
a
s a
cert
ain
l
o
cality, but it i
s
comp
romi
se
d
unde
r the limited com
put
ation and m
e
mory
re
so
urces. In the re
a
lization of FI
FO
strategy, ne
w entrants to the sam
p
le are put at the down
w
a
r
d an
d
t
he right colu
mn in matrix M,
and the elem
ents of the first ro
w and first colu
mn are
removed.
2.3. Semi-superv
ised Online Multiple Ker
n
e
l
Learning
Algorithm
for Big
Data
(SSOMK_bd)
Und
e
r th
e e
n
v
ironme
n
t of
big d
a
ta, lab
e
led
data lo
ss i
s
very
co
m
m
on a
nd th
e
big d
a
ta
set i
s
a mixe
d data
set in
cluding l
abel
e
d
data
and
u
n
label
ed d
a
ta
. If you only u
s
ed l
abel
ed
d
a
ta
to learn by
a
sup
e
rvised l
e
arnin
g
m
e
tho
d
, then
a
sup
e
rvise
d
le
arni
ng mo
del
doe
s n
o
t have
go
od
gene
rali
zatio
n
ability
a
n
d
it
can
cau
s
e
large
waste
of
unla
bele
d
data;
If
you only
u
s
e a
l
a
rge
amount
of un
labele
d
data
with impli
ed i
n
formatio
n to
learn by u
n
supervi
sed
lea
r
ning
metho
d
,
the unsupervi
sed learning
will ig
nore the value of labeled data.
S
e
mi-supervised learni
ng i
s
a
new m
a
chine
learni
ng met
hod bet
wee
n
the tradition
al sup
e
rvi
s
ed
learni
ng an
d
unsu
p
e
r
vise
d
learni
ng, its p
u
rpo
s
e i
s
to
make full
use
of a large
nu
mber
of unla
beled
sam
p
le
s to ma
ke up
for
the lack of labeled
sampl
e
s and imp
r
ov
e the learni
ng
performan
ce
effectively.
Figure 1. Semi-supe
rvise
d
Cla
ssifi
cati
on Flow
Cha
r
t
Algorithm 3. The Algorith
m
Descri
ptio
n of SSOMK_bd
Inpu
t
:
–D=
{
x
1
,x
2
,…,x
n
}
–Sampl
e
:(
x
t
,y
t
)
Outp
ut
:
Updated kern
el matri
x
K
Proc
e
dure
:
1) Initializ
ation
K
2) Learn K fro
m
D
0
3) for each (x
t
,y
t
) in D
4) if
y
t
is not
NULL the
n
5)
call SOMK_bd to car
r
y
out u
pdati
n
g
6
)
else
7
)
call UOMK_bd to carr
y out upd
ating
8
)
end if
9
)
end for
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Sem
i
-supe
rvi
s
ed O
n
line M
u
ltiple Kern
el Learning Alg
o
rithm
for Big Data (Ning Li
u)
643
Here, we pro
posed a sem
i
-su
p
e
r
vise
d online
multipl
e
kernel
l
earning algo
rith
m
for
bi
g
data whi
c
h i
s
nam
ed a
s
SSOMK_bd,
makin
g
full use of
samp
le labeling i
n
formatio
n a
n
d
unlab
eled
sa
mples with
i
m
plicit info
rm
ation to
ge
n
e
r
ate effe
ctive
ke
rnel
fun
c
tions an
d effe
ctive
cla
ssifie
r
. Specifically, wh
en re
ading
a sam
p
le, o
n
line lea
r
nin
g
method i
s
base
d
on t
he
increme
n
tal learni
ng to pe
rform. First it deter
mine
s
wheth
e
r the sampl
e
is ma
rke
d
. If marked,
SSOMK_bd use
s
su
pervi
sed
onlin
e multiple ke
rn
el
lea
r
nin
g
a
l
gorithm
SO
MK_bd pre
s
ented
above in
part
2.1 to ca
rry
out ke
rnel
m
odificatio
n
; otherwise, it uses the
on
-lin
e un
sup
e
rvised
multiple ke
rn
el learni
ng al
gorithm
UOM
K
_bd pr
esent
ed above in
part 2.2 to carry out kern
el
modificatio
n
s.
Iterative until
they meet certain
con
d
ition. Finally it
gene
rate
s th
e optimal
ke
rnel
function to form a cla
ssifie
r
and the test is ca
rri
ed out
on the test da
ta.
Among the
m
, online
multiple ke
rnel lear
ning
and se
mi-sup
ervised classificatio
n
algorith
m
a
r
e
sh
own in
Fi
gure
1,
uppe
r p
a
rt
within t
he d
a
shed
b
o
x rep
r
e
s
e
n
ts o
n
line
multi
p
le
kernel le
arni
ng an
d the
followin
g
pa
rt within the
dashed b
o
x rep
r
e
s
ent
s semi-sup
ervised
cla
ssifi
cation
pro
c
e
ss. At the end of the
online mu
ltip
le kernel lea
r
ning, it generates an o
p
timal
kernel fun
c
tio
n
for semi
-su
pervised cl
as
s
i
fic
a
tion to cons
truc
t the
class
i
fier.
The de
scripti
on of algorit
hm SSOMK_bd is sh
own
in Algorithm
3, where th
e kernel
function i
s
up
dated in ste
p
5) and 7
)
.
3. Experiment and Analy
s
is
The experi
m
ents are ca
rri
ed out on UCI data se
ts
and the pro
p
o
se
d algo
rith
ms in the
pape
r are
co
mpared with
the existing online mult
iple ke
rnel le
arnin
g
algo
rithm and batch
pro
c
e
ssi
ng of
multiple ke
rn
el learni
ng al
gorithm to ma
ke the effecti
v
eness a
s
sessment.
In the exp
e
ri
ment, SVM i
s
used
as a
ke
rnel
fun
c
tion.
RBF
ke
rnel
a
nd p
o
lynomia
l ke
rn
el
, whose para
m
eters are
selecte
d
ran
d
o
m
ly, are us
ed
to
c
o
ns
tr
uct k
e
r
n
e
l
fun
c
ti
on [23, 24]. And
0-1 lo
ss fun
c
t
i
on is u
s
ed to
evaluate the error rate.
In orde
r to
cancel the o
r
d
e
rs of mag
n
it
ude differen
c
e between th
e dimen
s
io
ns of data
and avoi
d large predi
ction
error
ca
used
by diffe
ren
c
es in i
nput a
nd outp
u
t, da
ta norm
a
lization
function
is u
s
ed h
e
re. As shown in
form
ula
(3),
the
i
n
put featu
r
e v
a
lue i
s
no
rma
lized
to [-1, 1]
by
data normali
zation functio
n
.
m
in
k
m
ax
()
/
(
x
x
)
kk
xx
x
(3)
Whe
r
e x
min
rep
r
e
s
ent
s the minimum value in data seq
uen
ce, x
max
repre
s
ent
s the maximum
value in data
seq
uen
ce.
Table 1. The
Dateset1 of Experiment
Index Dataset
Size
Dimensions
D1 Breast
683
9
D2 Splice
1000
60
D3 Dorothea
1150
100000
D4 Spambase
4601
57
D5 Mushrooms
8124
112
Table 2. The
Dateset2 of Experiment
Index Dataset
Size
Dimensions
D6 Forest
Cove
rT
yp
e
581012
54
D7 Poker-Hand
10
7
11
D8
Localization
Data for Pe
rson
Activity
164860
8
Such
a
s
Ta
bl
e 1
and
Ta
bl
e 2, 8
ki
nd
s
of
UCI data
sets a
r
e
use
d
in th
e exp
e
rime
nt.
Table 1 is u
s
ed to verify the validity of the SOMK
_bd
algorithm a
n
d
Table 2 is u
s
ed to verify the
validity of the
SSOMK_bd a
l
gorithm.
For the
samp
le sele
cted in
Table 2, the
prop
or
tio
n
of the trainin
g
set and the te
st set i
s
set to 1:1. The trainin
g
set divided into l
abeled a
n
d
unlabel
ed
sampl
e
s. Th
ree types of semi-
sup
e
rvised
cl
assificatio
n
e
x
perime
n
tal d
a
ta sets a
r
e
con
s
tru
c
ted
a
c
cordi
ng to t
he p
r
op
ortion
of
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 638 – 64
6
644
labele
d
samp
les to th
e nu
mber
of traini
ng set sampl
e
s. In first cl
a
ss, the
r
e
are
labele
d
samp
les
accou
n
ted fo
r 5% of the
sampl
e
s i
n
the traini
ng set; in the se
con
d
cla
s
s, there
are l
a
b
e
led
sampl
e
s a
ccounted for 1
0
% of the sample
s in the
training
set
;
in the third class, there
are
labele
d
sam
p
les a
cco
unte
d
for 15% of the sample
s in the training set. We ca
lculate the to
tal
cla
ssifi
cation
rate by cal
c
ul
ating the average
of the 3 kinds of cl
assif
i
cation rates.
At the same t
i
me, the data
in Table
2 is divided into
30 pa
rts, whi
c
h a
r
e to p
u
t in 4 file
s
sep
a
rately, a
nd the
data i
n
the
seq
uen
tial rea
d
file i
s
trai
ned
and
tested
on th
e onlin
e multi
p
le
kernel lea
r
ni
n
g
. All these d
a
ta are u
s
e
d
to
evaluate the rel
a
tionshi
p betwe
en th
e data set si
ze
and the CP
U pro
c
e
ssi
ng time.
Table 3. The
Experiment Result
OM-2
SOMK_bd
mistake(%)
time(s)
mistake(%)
time(s)
D1 34.30±2.80
0.270±0.015
24.25±3.10
0.245±0.013
D2 30.80±1.41
0.420±0.005
25.56±1.21
0.380±0.020
D3 10.70±0.72
0.435±0.013
8.20±0.52
0.426±0.041
D4 58.16±1.50
5.184±0.239
23.30±1.10
3.30±0.68
D5 0.37±0.03
8.691±0.070
0.34±0.02
2.560±0.02
Table 4. The
Corre
c
t Cla
s
sifi
c
a
tion Rat
e
Comparis
on (%)
Algorithm 1
Algorithm 2
SSOMK_bd
D6 72.5±1.80
74.25±2.10
77.0±2.43
D7 78.7±2.01
78.18±1.15
81.70±1.37
D8 61.02±0.24
67.55±1.33
75.46±1.65
Figure 2. CPU Ru
n Time
Comp
ari
s
ion
The first exp
e
rime
nt is ca
rrie
d
out to
compa
r
e the
prop
osed SO
MK_bd with
existing
online m
u
ltiple ke
rnel le
arning alg
o
rith
m om-2 [2
0]
to verify its
effec
t
ivenes
s
in reduc
i
ng
the
comp
utationa
l wo
rklo
ad
du
ring
ke
rnel
m
odificatio
n
. As sho
w
n in
T
able 3, the
ex
perim
ent results
sho
w
th
at SO
MK_bd
ha
s e
ffectiveness t
o
re
du
ce
ke
rnel
scale, ma
inly in sho
r
te
ning th
e lea
r
n
i
ng
time and red
u
cin
g
the error rate. The
rea
s
on i
s
tha
t
SOMK_bd redu
ce
s cal
c
u
l
ation wo
rkl
o
ad
durin
g ke
rn
el modificatio
n
and sele
cts some re
pre
s
e
n
tative kern
el
to update.
The
se
co
nd
e
x
perime
n
t is
carrie
d o
u
t to
co
mpa
r
e
the
propo
se
d
se
mi-supe
rvise
d
onli
n
e
multiple kern
el learni
ng al
gorithm SSO
MK_bd with
t
he existing
su
pervised onli
ne multiple
kerne
l
learni
ng
algo
rithm1 [21] a
n
d
alg
o
rithm2
[22] to ve
rif
y
it
s
ef
f
e
ct
iv
en
es
s
in
improving the
co
rre
c
t
cla
ssifi
cation.
As sho
w
n in
Table
4, we
c
an
kn
ow th
at the propo
sed SSOMK_
bd ha
s hi
ghe
r
corre
c
t cla
s
si
fication rate
than the existing alg
o
rith
m1 [21] and
algorithm
2 [22]. Espe
cial
ly,
SSOMK_bd i
s
effe
ctive in
dealin
g with
l
a
rge
scal
e in
compl
e
te la
b
e
led
data. Th
e re
ason i
s
t
hat
SSOMK_bd
make
s full u
s
e of the tag i
n
formatio
n of
labele
d
sam
p
les
and
abu
ndant info
rm
ation
of unlab
eled
sam
p
le
s to
train effe
ctive cla
ssifie
r
,
whi
c
h
can
improve th
e
perfo
rma
n
ce of
learni
ng.
0
5
10
15
20
25
30
0
5
10
15
20
25
Run t
i
m
e
v
a
r
i
at
i
o
n
di
ag
ram
T
h
e s
c
al
e
of
da
t
a
s
e
t
(
*
1
0
6
)
T
he run t
i
m
e
of
C
P
U
/
s
(*
10
2
)
Al
g
o
r
i
t
h
m
1
Al
g
o
r
i
t
h
m
2
SS
O
M
K
b
d
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Sem
i
-supe
rvi
s
ed O
n
line M
u
ltiple Kern
el Learning Alg
o
rithm
for Big Data (Ning Li
u)
645
The data in T
able 2 was di
vided into 30
par
ts to eval
uate the rel
a
tionship bet
ween the
data set si
ze
and the
CPU
pro
c
e
ssi
ng ti
me. As sho
w
n in Figu
re
2, we
can
se
e
SSOMK_bd
has
a goo
d lin
ear relatio
n
ship
betwe
en the
gro
w
th
rate o
f
the ru
nnin
g
time an
d dat
a set si
ze. T
h
e
experim
ent result in
dicate
s that SSOM
K
_bd is
with
good
scalabili
ty and co
uld
be u
s
ed in
m
o
re
regul
ation
s
d
a
ta analysi
s
and ap
plication. T
he rea
s
on is that S
S
OMK_bd u
s
e
s
increm
e
n
tal
learni
ng f
r
am
ewo
r
k to im
p
r
ove the
cl
assificatio
n
p
e
rf
orma
nce a
n
d
it is
effective
in the
big
da
ta
environ
ment.
4. Conclusio
n
In this pa
per,
the online
mu
ltiple ke
rnel l
e
ar
nin
g
alg
o
rit
h
m und
er bi
g
data environ
ment is
studie
d
deepl
y, and the semi-supe
rvise
d
learni
ng is
i
n
trodu
ce
d into the field of
big data ma
chin
e
learni
ng. Th
e
tradition
al kernel
lea
r
nin
g
algo
rithm
s
is imp
r
oved
to red
u
ce t
he comp
utational
worklo
ad du
ri
ng ke
rnel mo
dification. Th
e increme
n
tal learnin
g
fra
m
ewo
r
k is used to improve
the
cla
ssifi
cation
perfo
rman
ce
in big data e
n
vironm
ent. Based o
n
the
current readi
ng of larg
e d
a
ta
fragme
n
ts i
n
an o
n
line
wa
y, the algo
rith
m ma
ke
s full
use
of the
ri
ch information
of lab
e
led
d
a
ta
and unl
abel
e
d
data to a
c
hieve ke
rn
el
updating
an
d
co
nstruct
efficient kern
el function.
The
experim
ent is condu
cted
on the be
n
c
hma
r
k UCI
large
data set, the results show th
at th
e
prop
osed
alg
o
rithm
s
a
r
e e
ffective in sh
ortenin
g
the l
earni
ng time
and
red
u
ci
ng
the erro
r rate.
Also the prop
ose
d
algo
rith
ms could be
use
d
in big d
a
ta analysi
s
a
nd appli
c
atio
n.
Ackno
w
l
e
d
g
ements
The Proje
c
t Suppo
rted
by Natural Sci
e
n
c
e Ba
si
c Research Pla
n
in
Shaanxi Pro
v
ince o
f
Chin
a
(Prog
r
am
No.20
1
5
J
M63
4
7
)
; Th
e Proj
ect
Su
pporte
d
by
Scien
c
e
Te
chnolo
g
y Plan
in
Shanglu
o
Cit
y
of China (Program No. SK2014-
0
1
-15); The Project Supp
orted by Scien
c
e
Research Plan of Shangluo Univ
ersity (Program No.14SKY026).
Ref
e
ren
c
e
[1]
LI W
u
jun, Z
h
o
u
Z
h
ihu
a
. Lear
nin
g
to hash fo
r big d
a
ta: Current status a
nd future tren
ds.
Chin
ese
Scienc
e Bull
eti
n
. 2015; 6
0
(5-6
): 485-49
0.
[2]
Ma
yer-Sch
önb
erger V, Cuki
e
r
K. Big Data:
A
Revol
u
tion T
hat Will T
r
ansform Ho
w
We Live, Work,
and T
h
ink.
America
n
Journ
a
l
of Epide
m
iol
o
g
y
. 2014; 17(
1): 181-
183.
[3]
Li Z
h
i
jie,
Li
Y
uan
xian
g, W
a
ng F
e
ng, et
al
. Onlin
e L
earn
i
ng
Alg
o
rithms
for Big
Dat
a
Anal
ytics: A
Survey
.
Journ
a
l
of Computer
Rese
arch an
d
Devel
o
p
m
ent
. 201
5; 52(8): 17
07-1
721.
[4]
Z
hou Z
H
, Cha
w
l
a
NV, Ji
n Y, et al. Big Dat
a
Opportu
nitie
s
and C
hal
le
n
ges: Discuss
io
ns from Dat
a
Analy
t
ics Pers
pectives.
IEEE Co
mp
utation
a
l
Intelli
genc
e Ma
ga
z
i
n
e
. 20
14; 9(4): 62-7
4
.
[5]
T
an M,
T
s
ang IW
, W
ang L. T
o
w
a
rds
Ultrah
i
g
h
Dim
ensi
o
n
a
l
F
eature S
e
lecti
on for B
i
g
Data
.
Journa
l o
f
Machi
ne Le
arn
i
ng R
e
searc
h
. 201
4; 15(2): 13
71-1
429.
[6]
Yoo C, Ram
i
re
z L, Liuzzi J. Bi
g data a
n
a
l
ysis
using m
oder
n
statistical an
d
machi
ne l
earn
i
ng meth
ods
i
n
me
di
ci
ne
.
Internati
ona
l Ne
u
r
ourol
ogy Jo
ur
nal
. 20
14; 18(
2
)
: 50-57.
[7]
HE Qing, LI N
i
ng, L
U
O W
e
n
-
Juan, et
al. A
Surv
e
y
of Ma
chin
e Le
arn
i
ng
Algor
ithms for
Big D
a
ta.
Pattern Reco
g
n
itio
n & Ar
tificial Intelligence
.
201
4; 27(4): 32
7-33
6.
[8]
Hans
en T
J
,
Maho
ne
y M W
.
Semi-su
pervis
ed Eig
env
ector
s
for Large-sc
ale L
o
cal
l
y
-bi
a
sed Le
arni
ng.
Journ
a
l of Mac
h
in
e Le
arni
ng
Rese
arch
. 20
1
4
;15(1
1
): 369
1
-
373
4.
[9]
Klei
ner A, T
a
l
w
alkar A, S
a
rkar
P, et al.
T
h
e b
i
g d
a
ta
bootstr
a
p
. Proc
eed
in
g
s
of the
29th
In
ternatio
nal
Confer
ence
on
Machin
e Le
arnin
g
(ICML). Edin
burg
h
. 201
2
:
1759-1
7
6
6
.
[10]
Gonzal
ez JE,
Lo
w
Y, Gu
H, et al.
Pow
e
r
G
raph: Distri
b
uted gr
ap
h-par
alle
l co
mputati
on o
n
n
a
tura
l
grap
hs
. Proc
eed
ings
of t
he
10th
US
ENIX S
y
m
pos
ium
on O
per
ating
S
y
stem
s Des
i
gn
an
d
Impleme
n
tatio
n
(OSDI). Holl
yw
o
o
d
. 201
2: 17-30.
[11]
Gao W
,
Jin R
,
Z
hu S, et al.
One-pass AUC opti
m
i
z
at
io
n
. Proceed
ing
s
of the 30th Internatio
na
l
Confer
ence
on
Machin
e Le
arnin
g
(ICM
L). Atlanta. 20
13: 90
6-91
4.
[12]
Hoi S
CH, W
a
ng J, Z
h
a
o
P. LIBOL: A Libr
ar
y
f
o
r
Onli
ne Lear
nin
g
Al
gor
ithms.
Journ
a
l
of Machi
n
e
Lear
nin
g
Res
e
arch
. 201
4; 15(
1): 495-4
99.
[13]
Yang
Ha
ifen
g, Liu
Y
u
a
n
, Xie
Z
henp
ing,
et a
l
. Efficientl
y
T
r
ainin
g
B
a
ll
Vect
or Mach
in
e i
n
Onlin
e W
a
y.
Journ
a
l of Co
mputer Res
earc
h
and D
e
ve
lop
m
e
n
t
. 2013; 5
0
(
9): 1836-
18
42
.
[14]
Shal
ev-Sh
w
a
r
tz S, Z
hang T
.
Acceler
a
ted pr
oxi
m
al stoc
has
tic dual
c
oord
i
nate asc
ent for
regul
ari
z
e
d
loss minimi
z
a
tion
. Procee
di
n
g
s of the 31st
Internation
a
l
Confer
ence
on
Machin
e Le
ar
nin
g
(ICML).
Beiji
ng. 20
14:
64-7
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 14, No. 2, June 20
16 : 638 – 64
6
646
[15]
Yang
H, F
o
ng
S.
In
crem
en
tal
l
y
op
tim
i
z
e
d
de
ci
si
on
tree
for n
o
i
sy bi
g
d
a
t
a
. Proce
edi
ng
s of the
1st
Internatio
na
l W
o
rkshop o
n
Big Data, Streams an
d H
e
terog
e
n
eous
Source Mi
nin
g
:
Algorithms,
S
y
stems, Prog
ramming Mo
de
ls and Ap
plic
ati
ons. ACM. 201
2.
[16]
Motai Y. K
e
rn
el Ass
o
ciati
o
n
for Cl
assificati
on
an
d Pre
d
ict
i
on: A
Surv
e
y
.
IEEE
Trans
Neur
al Netw
Lear
n Syst
. 2014; 26(2): 2
08-
223.
[17]
Z
heng
H, Ye
Q, Jin Z
.
A Novel Multi
p
l
e
Ke
rnel
S
parse
Re
prese
n
tatio
n
b
a
sed
Class
ifica
t
ion for F
a
ce
Reco
gniti
on.
K
s
ii T
r
ansacti
on
s on Internet & Information Sy
stems
. 20
14; 8
(
4): 1463-
14
80
.
[18]
Shrivastav
a A,
Pill
ai JK,
Pat
e
l VM. Mu
ltipl
e
ke
r
nel-
base
d
dictio
nar
y
le
ar
nin
g
for
w
e
ak
l
y
s
u
p
e
rvise
d
classification.
Pattern
Reco
g
n
itio
n
. 201
5; 48: 2667-
26
75.
[19]
Liu K
H
, L
i
n
YY, Che
n
CS
. Lin
ear S
pec
tral
Mi
xture
A
nal
ysis
vi
a M
u
ltipl
e
-Ker
nel
Lear
nin
g
for
H
y
pers
pectral
Image
C
l
ass
i
fi
cation.
IEEE T
r
ansactions
on Geo
science
& Rem
o
t
e
Sensing
. 20
15;
53(4): 22
54-
22
69.
[20]
Luo
J, Orab
on
a F
,
F
o
rno
n
i
M, et al.
OM-2
: An Onl
i
ne
M
u
lti-class
Multi-
kerne
l
L
ear
nin
g
Al
gorith
m
.
Procee
din
g
of CVPR 20
10. 2
010: 43-
50.
[21]
Orabon
a F
,
L
uo J, C
a
p
u
to
B. Multi K
e
r
nel
Le
arni
ng
w
i
t
h
o
n
li
ne-
bat
ch o
p
timizati
o
n
.
Jour
nal
of
Machi
ne Le
arn
i
ng R
e
searc
h
. 201
2; 13(1): 22
7-25
3.
[22]
Jin R,
Ho
i SC
H, Yan
g
T
.
Online
Mu
ltipl
e
K
e
rne
l
L
ear
nin
g
: Alg
o
rithms
an
d Mistak
e Bo
u
nds.
Le
ctu
r
e
Notes in C
o
mp
uter Scienc
e
. 2
010; 63
31(
4): 390-4
04.
[23]
Z
hang
W
,
Guo
Z
,
Z
hang
L,
et al. A
d
a
p
tive
Co
ntrol f
o
r R
obotic
Man
i
p
u
l
a
tors b
a
se
on
RBF
Ne
ura
l
Net
w
ork.
TEL
K
OMNIKA Teleco
mmu
n
icati
o
n Co
mp
utin
g
Electron
ics an
d Co
ntrol
. 20
1
3
; 8(3): 49
7-
502.
[24] Pan LJ, Ji
n
W
,
W
u
J. A Novel I
n
trusi
on D
e
tectio
n Appro
a
ch
usin
g Multi-K
e
rne
l
F
unctions.
T
E
LKOMNIKA T
e
leco
mmunic
a
tion C
o
mputi
n
g Electron
ics a
nd Co
ntrol
. 20
14; 12(4): 1
088
-109
5.
Evaluation Warning : The document was created with Spire.PDF for Python.