Indonesi
an
Journa
l
of
El
ect
ri
cal Engineer
ing
an
d
Comp
ut
er
Scie
nce
Vo
l.
12
,
No.
3
,
Decem
ber
201
8
, p
p.
1
0
63~
1
0
70
IS
S
N: 25
02
-
4752, DO
I: 10
.11
591/ijeecs
.v1
2
.i
3
.pp
1
0
63
-
1
0
70
1063
Journ
al h
om
e
page
:
http:
//
ia
es
core.c
om/j
ourn
als/i
ndex.
ph
p/ij
eecs
Adaptiv
e Data S
tructu
re
Based O
versamp
lin
g A
l
gorithm
f
or
Ordinal
Classific
atio
n
D.
Dhan
alaks
hmi
1
,
A
nn
a
S
ar
o Vijen
dran
2
1
Depa
rtment
of
Com
pute
r
Scie
n
ce
,
Sri R
amakri
shna
Col
le
ge
of A
rts
and
Sci
ence
,
Coim
bat
or
e, I
n
dia
2
School
of
Com
puti
ng,
Sri R
amakri
shna
Col
le
ge
of
Arts a
nd
Science
,
Coim
bat
or
e
,
Indi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ma
y
1
, 2
01
8
Re
vised
Jun
25
, 201
8
Accepte
d
Aug
2
1
, 201
8
The
m
ai
n
obje
ctive
of
thi
s
rese
a
rch
is
to
improv
e
the
pre
dictiv
e
ac
cur
acy
of
cl
assifi
ca
t
ion
in
ordina
l
m
ulticlass
imbala
nce
d
sce
nar
io
.
Th
e
m
et
hodol
o
g
y
at
t
empts
to
upli
ft
th
e
c
la
ss
ifi
er
p
erf
orm
an
ce
through
s
ynthe
sizi
n
g
sophistic
at
ed
ob
je
c
ts
of
imm
at
ure
class
es.
A
novel
Adapt
ive
Dat
a
Struct
ur
e
base
d
Oversam
pli
ng
al
gori
thm
i
s
proposed
to
c
rea
t
e
s
y
ntheti
c
obje
c
ts
and
Ext
reme
Learni
ng
Mac
hine
for
Ordina
l
Regre
ss
ion
(EL
MO
P)
cl
assifi
er
is
adopt
ed
to
v
alid
at
e
our
work.
T
he
proposed
m
ethod
gene
ra
ti
ng
new
obje
c
ts
b
y
an
al
y
z
ing
th
e
cha
ra
ct
er
isti
cs
and
int
ri
cac
y
of
imm
at
ure
cl
ass
obje
c
ts.
On
the
whol
e,
the
d
at
a
set
is
div
ide
d
int
o
training
an
d
te
st
data.
Tr
ai
n
ing
da
ta
s
et
is
updated
wit
h
new
s
y
ntheti
c
obj
ec
ts.
Th
e
expe
r
imenta
l
ana
l
y
sis
is
per
form
ed
on
t
esti
ng
da
ta
s
et
to
ch
e
ck
the
eff
iciency
of
th
e
proposed
m
et
hodolog
y
b
y
compari
ng
it
with
the
exi
st
in
g
work.
Th
e
p
erf
orm
anc
e
eva
lu
at
ion
is
co
nduct
ed
in
te
rm
s
of
the
par
amete
rs
ca
lled
Mea
n
Abs
olut
e
Err
or,
Maximum
Mea
n
Ab
solute
Err
or,
Geo
m
et
ric
Me
an,
Kappa
and
Avera
ge
Acc
ur
a
c
y
.
Th
e
m
ea
sure
s
prove
tha
t
the
proposed
m
et
hodolog
y
c
an
produc
e
aut
h
en
ti
c
s
y
n
the
t
ic
o
bje
c
t
s
tha
n
the
exi
sting
te
chn
ique
s.
The
Propos
ed
te
chn
ique
c
an
s
y
n
th
esiz
e
the
new
eff
ective
objects
through
eva
lu
at
ing
t
he
st
ruc
ture
of
imm
at
ure
cl
ass.
It
boo
sts
the
glob
al
pr
ec
ision
and
cl
ass wise
p
recision
espe
ci
a
lly
pr
ese
rve
s r
ank
ord
er
of
the classes.
Ke
yw
or
d
s
:
Ad
a
ptive
data
structu
r
Av
e
ra
ge
acc
uracy
Extrem
e lea
rn
ing m
achine for
ordinal
regress
ion
Geo
m
et
ric
m
ea
n
Kappa
Ma
xim
u
m
m
ea
n
a
bsolute er
r
or
Mult
i cl
ass o
r
di
nal
cl
assifi
cat
ion
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
D.
D
han
al
a
ksh
m
i
,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce,
Sr
i R
am
akr
ish
na
C
ollege
of
Ar
ts a
nd Scie
nc
e
,
Coim
bator
e, In
dia.
Em
a
il
:
dh
ana
duraira
j@gm
ai
l.
com
1.
INTROD
U
CTION
1
.
1.
B
ackgro
und
Ty
pical
cl
assif
ic
at
ion
al
gorit
hm
s
well
beh
a
ve
d
with
a
ppr
op
riat
el
y
balance
d
dataset
s,
but
m
any
real
-
world
a
ppli
cat
ion
s
i
n
va
rio
us
discipli
nes
e
xh
i
bit
i
m
balanced
or
din
al
na
ture
s
uch
a
s
Disease
pre
di
ct
ion
,
Weathe
r
f
or
ec
ast
ing
,
Pe
rfo
r
m
ance
pr
e
dicti
on,
Ra
ti
ng
a
nd
Finan
ci
al
in
ve
st
m
ent
et
c
.,
Most
of
the
cl
ass
ific
at
ion
al
gorithm
s
try
to
achie
ve
th
e
bette
r
perform
ance
globall
y.
They
s
uffer
t
o
ob
ta
in
bette
r
lo
cal
as
well
as
global
perform
ance
due
to
s
ke
we
d
ordinal
nat
ure
of
data.
Th
e
m
ai
n
resear
ch
co
ntributi
ons
in
this
perspecti
ve
include
Algori
thm
ic
le
vel,
da
ta
le
vel
a
nd
cost
sensiti
ve
appr
oach
es
[
1]
,
[2
]
Pr
op
os
e
d
var
ia
ti
on
of
sm
ot
e
al
gorithm
SN
OCC
for
i
m
balanced
bin
a
ry
cl
ass
wh
ic
h
rec
ognizes
m
or
e
than
tw
o
see
d
sa
m
ples
to
create
new
sam
ples
in
the
interi
or
re
gion
of
the
bor
de
red
see
d
sam
ples
to
sim
ulate
the
e
ven
an
d
uneve
n
distri
bu
t
ion
of
or
i
gin
al
sam
ples
[
3]
.
Test
e
d
var
i
ou
s
cl
assifi
ers
perform
ance
base
d
on
c
ost
for
tw
o
cl
as
s
i
m
balanced
public
healt
h
dataset
pro
blem
.
They
con
cl
uded
t
ha
t
Ba
ye
sia
n
cl
assif
ie
rs
work
well
for
this
pro
blem
[4]
.
Propose
d
On
li
ne
ve
rsion
of
Im
balanced
S
VM
(OIS
V
M)
for
bin
ary
e
m
ai
l
c
la
ssi
ficat
ion
to
im
pr
ove
pr
ocessi
ng
sp
ee
d
and
sa
ve
stora
ge
sp
ace
[
5]
.
P
rove
d
ELM
is
eff
ic
ie
nt
al
gori
thm
fo
r
cl
assifi
cat
ion
pro
blem
[6
]
.
Confirm
e
d
that,
Eucli
dea
n
an
d
Ci
ty
blo
ck
distance
m
eas
ur
es
perf
or
m
e
d
well
in
K
-
Near
est
Neig
hbou
r
Algo
rith
m
[
7
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
1
0
63
–
1
0
70
1064
Pr
op
os
e
d
firs
t
ov
e
rsam
pling
al
gorithm
to
han
dle
im
balanced
m
ulti
c
la
ss
ordinal
cl
assifi
cat
ion
pr
ob
le
m
[8]
.
Au
t
hors
pe
rform
ed
li
te
ratur
e
rev
ie
w
in
the
con
te
xt
of
ordinal
cl
assifi
cat
i
on
to
fi
nd
th
e
causes
f
or
cl
as
sifie
r
perf
orm
ance
de
gr
a
datio
n
[
9]
,
[10]
.
A
uthors
pro
po
se
d
propose
d
colli
near
base
d
oversam
pling
al
gorith
m
in
the
safe
an
d
bor
de
r
li
ne
re
gion
for
ordi
nal
cl
assifi
cat
ion
[
11
]
.
I
ntr
oduce
d
uns
up
e
r
vised
over
sam
pling
m
et
ho
d
f
or
ordinal
re
gr
es
s
ion
[
12
]
.
Ado
pt
data
char
act
ei
sti
cs
to
identify
co
m
plex
obje
ct
s
and
deci
de
si
ze
to
synthesiz
e
for
su
c
h
eac
h
pro
blem
atic
fo
r
le
ar
ni
ng
as
well
as
m
os
t
resp
on
sible
for
pe
rf
or
m
ance
degr
adati
on
obj
ect
s
[
13
]
A
dopted
cl
us
te
ri
ng
t
o
group
m
inorit
y
instanc
es
an
d
synt
hes
iz
ing
ob
j
ect
s
ba
sed
on
t
he
le
arn
i
ng
com
plexity
of
the
gro
up
[
14
]
Ma
ke
us
e
of
data
char
act
e
risti
cs
fo
r
ove
rs
a
m
pling
ob
j
ect
s
and
c
on
cl
ud
e
d
tha
t
insig
ht
the
f
orm
at
ion
and
gr
oup
of
obj
ect
s
in
dataset
coll
us
io
n
the
pr
e
possessi
ng
al
gor
it
h
m
s
wh
ic
h
upli
ft
the
perform
ance.
Au
t
hors
[
15
]
s
uggeste
d
that,
synthesiz
i
ng
m
or
e
obj
ect
s
i
n
the
b
orde
rline
wh
ic
h
e
xpa
nd
the
pro
bab
le
a
rea
of im
m
at
ur
e cla
ss.
1
.
2
.
Pr
ob
le
m
Ind
en
tified
f
r
om
Li
ter
ature
R
e
view
Most
of
t
he
e
xi
sti
ng
wor
k
f
oc
us
es
on
im
balanced
bin
ary
c
la
ss
cl
assifi
cat
ion,
im
balanced
m
ulti
cl
ass
cl
assifi
cat
ion
a
nd
or
din
al
cl
as
sific
at
ion
al
one.
Ver
y
fe
w
re
search
w
orks
ha
s
bee
n
ca
rr
ie
d
out
f
or
im
bala
nc
e
d
m
ul
t
ic
la
ss
or
di
nal
cl
assifi
cat
ion.
He
re
we
de
al
su
ch
a
c
om
pl
ic
at
ed
sit
uation
Im
balanc
ed
Mult
ic
la
ss
Ordinal
Cl
assifi
cat
ion
.
Ma
ny
real
ti
m
e
app
li
cat
ions
in
var
io
us
dom
ai
ns
su
ch
a
s
Eco
nom
y,
A
uto
m
ob
il
e
I
ndus
try
,
Me
dicine,
A
gri
culture
,
Bi
oM
edici
ne,
Hum
an
Re
so
urce
Dev
el
op
m
ent
et
c.,
co
ns
ist
s
data
in
im
ba
la
n
ce
d
m
ul
ti
cl
ass
or
di
nal
nat
ur
e.
T
hey
dem
and
eff
ect
ive
sta
te
-
of
-
t
he
-
a
rt
s
olu
ti
on
to
ta
ckl
e
this
sce
nar
i
o
for
i
m
pr
ovem
ent o
f pr
e
dicti
ve
ac
cur
acy
a
nd m
i
nim
iz
at
ion
of e
rror rate.
1
.
3
.
Pr
opose
d Soluti
on
In
this
pa
per
,
we
pro
pose
an
A
da
ptive
Data
Stru
ct
ur
e
Ba
sed
O
ver
s
a
m
pling
Al
gor
it
h
m
.
In
ou
r
pro
po
se
d
a
dapt
ive
data
str
uctur
e
base
d
ove
r
sam
pling
al
gor
it
h
m
diff
ers
from
the
existi
ng
w
ork
[
7
]
it
prefe
rs
o
nly
the
com
plica
te
d
ob
j
ec
ts,
com
par
e
with
[
12
]
it
handles
m
ulticlass
ordi
nal
cl
assifi
cat
ion
,
dev
ia
te
d
with
[
14
]
our
a
lgorit
hm
han
dl
es
m
ult
ic
la
ss
or
di
nal
cl
assifi
cat
ion
a
nd
w
ork
s
on
each
patte
rn
s
of
m
ino
rity
cl
ass
to an
al
yse
s c
om
plexit
y.
2.
ADAPTI
VE
D
AT
A
S
T
R
U
CTU
RE
B
A
S
ED
O
VER
SAMPLI
NG
A
L
GORIT
HM
2
.
1.
Sele
c
ting Imma
tu
re
C
l
as
s
The
dataset
is
par
ti
ti
on
e
d
in
to
trai
ning
an
d
te
sti
ng
gro
up.
[
7]
,
[
16
]
,
[
17
]
Dif
fer
e
nt
works
in
th
e
li
te
ratur
e
hav
e
consi
der
e
d
t
he im
m
at
ur
e cla
ss
es that e
xh
i
bit IR v
al
ue
a
bove
than 1
.5.
q
q
j
j
q
N
.
Q
N
IR
(1)
In
this
w
ork
usi
ng
(
1),
IR
va
lued
is
cal
cula
te
d
f
or
eac
h
cl
ass.
Ba
se
d
on
that
val
ue,
t
he
num
ber
of
syntheti
c
patte
rn
s
t
o
be
ge
ne
rated
for
eac
h
cl
ass
is
cal
c
ulate
d
th
rou
gh
(2).
Af
te
r
(
2),
agai
n
IR
value
is
cal
culat
ed
for
each
cl
ass.
Wh
e
n
(
2)
al
te
rs
th
e
IR
val
ue
for
t
he
re
m
ai
nin
g
cl
ass
es
as
a
bove
t
han
1.5,
new synt
hetic
o
bject
s a
re c
re
at
ed
f
or s
uch cla
sses unti
l reac
hes
t
he
IR
n
e
w
(3)
value
b
el
ow
1.5.
q
q
j
q
Q
c
c
c
j
q
N
Q
Th
r
e
s
h
o
l
d
S
y
n
S
y
n
N
S
y
n
.
1
(2)
Q
Sy
n
N
Sy
n
Sy
n
N
ne
w
IR
q
q
q
j
q
Q
c
c
c
j
q
.
)
(
1
(3)
2
.
2
.
Adapt
in
g Data S
truc
t
ure t
o Overs
am
ple
In
im
m
at
ur
e
cl
ass,
5
Near
e
st
Neig
hbors
for
each
obj
ect
is
cal
culat
ed
a
nd
cl
assify
eac
h
obj
ect
i
nto
secur
e
ob
j
ect
s
and insec
ur
e
obj
ect
s
b
ase
d o
n nea
rest n
ei
ghbors
.
K
r
j
j
,
s
C
Im
.
.
.
1
j
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ad
ap
ti
ve
Da
t
a Struct
ur
e B
as
e
d Overs
amplin
g
Al
go
rit
hm f
or Or
dinal Cl
assi
fi
cation
(
D.D
hanalaks
hmi
)
1065
Am
on
g
5
Nea
rest
Neig
hbors
,
4
or
5
neig
hbors
belo
ng
to
i
m
m
a
ture
cl
ass
it
con
side
red
as
sa
fe
obj
ect
s,
2
or
3
m
eans
it
is
cal
le
d
as
bor
der
l
ine
ob
j
ect
,
1
m
eans
it
is
ra
r
e
obj
ect
an
d
f
or
0
it
is
outl
ie
r
[
14
]
.
In
our
propos
ed
w
or
k
we
tr
eat
ed
a
bove
s
ai
d
obj
ect
cat
egories
i
nto
2
gro
ups
s
uc
h
as
sec
ur
e
ob
j
e
ct
s
an
d
insecu
re
ob
j
ec
ts.
Safe
ob
j
ect
is
secur
e
obj
e
ct
re
m
ai
nin
g
c
at
egorie
s
belo
ng
to
i
ns
ecu
re
obj
ect
s.
F
or
ordi
nal
cl
assifi
cat
ion
s
cenari
o,
a
dj
ace
nt
cl
asses
are
c
lose
to
eac
h
ot
h
er
[7
]
.
C
onsid
erati
on
of
the
a
bove
sta
te
m
ent
,
f
or
the
sec
ur
e
ob
j
e
ct
,
1
nea
rest
ne
ighbor
of
the
non
-
im
m
at
ur
e
c
la
ss
obje
ct
bel
ong
t
o
one
of
the
a
djacent
cl
a
ss
that
safe
obj
ect
s
is
inten
de
d
as
ordi
nal
bo
rd
e
rline
obj
ect
.
O
r
di
nal
bor
de
rline
obje
ct
s
a
nd
insec
ur
e
ob
j
ect
s
ar
e
captu
red f
or fu
rther p
r
ocessin
g.
The
n
e
w
im
m
at
ur
e cla
ss (I
m
C) co
nsi
sts t
he
a
bove sa
id t
wo grou
ps
.
Im
C=
{Or
din
al
bor
der
li
ne o
bje
ct
s,
insec
ur
e
obj
ect
s}
(5)
Im
C=
j
x
x
x
.
.
.
,
2
1
(6)
2
.
3.
Fin
ding
Ad
j
acen
t Clas
ses and
Synth
esi
z
ing O
bj
ec
t
s
Adjace
nt
cl
ass
patte
rn
s
are
ve
ry
cl
os
e
to
i
m
m
a
ture
cl
ass
patte
rn
s
com
par
e
with
nona
dj
ace
nt
cl
ass
patte
rn
s
[7
]
.
A
ccordin
g
to
tha
t,
this
wo
r
k
fin
ds
the
shortest
distance
for
e
ach
obj
ect
of
both
ad
j
ace
nt
cl
asses
thr
ough im
m
atu
re
class
obj
ec
ts.
2
.
4
.
Gr
aph
C
on
s
truct
i
on
Creat
e
gr
a
ph
f
or
t
he
im
m
at
ur
e
cl
ass,
q
be
the
in
dex
of
the
cl
ass
we
w
ant
to
ov
e
r
-
sa
m
ple.
Creat
e
gr
a
ph
q
G
for
cl
ass
q
C
Im
ba
sed
on t
hr
e
e sub
gr
a
phs
q
q
G
,
1
,
q
q
G
,
a
nd
1
,
q
q
G
a.
Con
st
ru
ct
q
q
G
,
1
Fo
r
e
ve
ry
patte
rn
in
q
th
cl
ass,
fin
d
it
s
k
-
near
e
st
neig
hbor
in
the
q
-
1
th
cl
ass
us
i
ng
the
f
or
m
ul
a
)
,
,
(
1
k
X
X
N
q
q
d
.
Creat
e e
dges.
Fo
r
e
ve
ry
patte
rn
in
q
-
1
th
cl
ass,
fin
d
it
s
k
-
near
est
neighb
or
in
t
he
q
th
cl
ass
us
in
g
the
f
or
m
ul
a
)
,
,
(
1
k
X
X
N
q
q
d
.
Creat
e e
dges.
b.
Con
st
ru
ct
gr
a
ph
q
q
G
,
1
with e
dg
e
s
only
those are
c
omm
on
in
:
)
,
,
(
1
k
X
X
N
q
q
d
)
,
,
(
1
k
X
X
N
q
q
d
c.
Co
ns
t
ru
ct
q
q
G
,
Fo
r
e
ver
y
patt
ern
in
q
th
cl
ass
,
find
it
s
k
nea
rest
neig
hbour
s
in
the
q
th
cl
ass
and
create
e
dg
e
s
with
thes
e
neig
hbours
.
d.
Con
st
ru
ct
1
,
q
q
G
sam
e li
ke
q
q
G
,
1
e.
Find
t
he
s
hortest
path
from
q
q
G
,
1
to
1
,
q
q
G
via
q
q
G
,
us
in
g
Dijkstra
’s
al
gorithm
fo
r
eac
h
ve
rtex
i
n
q
q
G
,
1
f.
Sele
ct
an
e
dge
from
, an
d
base
d on ove
rsam
pling
rate t
hat s
hould be
one
of the
sho
rtest
pat
h
ed
ge
.
3.
FUR
T
HER
CON
CER
NS
To
cl
arify
al
l
the
w
orks
wh
i
ch
are
done
i
n
the
pr
e
vious
subsect
ion,
a
su
m
m
ary
of
the
w
ork
is
giv
e
n
:
3.1.
Pse
ud
o
Code
f
or
t
he
Prop
os
ed
A
D
SOS
Inp
ut
: Traini
ng D
at
aset
Ou
t
pu
t
:
New B
al
anced
Trai
ne
d datase
t
Pha
se
I
1)
Sele
ct
the im
mature class
to b
e oversam
pled (im
)
based
on IR
v
al
ue
, calcul
at
ed
usi
ng e
quat
ion
(1).
2)
The n
um
ber
of
obj
ect
s
to be s
ynthesiz
ed
is c
al
culat
ed usin
g eq
uation (
2).
3)
The
new
IR
va
lue
is
cal
culat
ed
usi
ng
(
3).
U
nt
il
IR
value
f
or
al
l
the
cl
asses
reach
le
ss
t
han
1.5
re
pea
t
ste
p b to st
ep
c
.
Pha
se
II
Find the
struct
ur
e
of the
im
mature class
u
si
ng
e
qu
at
io
n (
4)
Secu
re
Ordinal
obj
ect
s
and i
nse
cur
e
ob
j
ect
s
are
deliberate
d f
or
oversam
pling.
Pha
se
III
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
1
0
63
–
1
0
70
1066
Con
st
ru
ct
gr
a
ph
q
q
G
,
1
,
q
q
G
,
an
d
1
,
q
q
G
as m
entioned in t
he
a
bove se
ct
ion
Find
sho
rtest
pat
h
f
ro
m
q
q
G
,
1
to
1
,
q
q
G
via
q
q
G
,
Ra
ndom
ly
sele
ct
an
e
dg
e
from
q
q
G
,
1
,
1
,
q
q
G
an
d
q
q
G
,
Pha
se
IV
Synthesizi
ng
ne
w objects
Sele
ct
ed
e
dg
e
belo
ngs to
q
q
G
,
app
l
y un
i
form
d
ist
ribu
ti
on
for
sy
nt
hesizi
ng ob
j
ec
ts
Sele
ct
ed
e
dg
e
belo
ngs to
q
q
G
,
1
or
1
,
q
q
G
app
li
es
gam
m
a
distrib
ution f
or synthesizi
ng
obj
ect
s
Pha
se
V
Pr
e
dicti
on
us
in
g ordinal cl
assi
fier
Inp
ut
: Ne
w
Ba
la
nced
Trai
ned D
at
aset
, Te
st
dataset
Ou
t
pu
t:
P
re
dicte
d value
4.
RESU
LT
S
A
ND AN
ALYSIS
To
validat
e
the
pro
posed
m
eth
od
ology
s
ome
data
set
s
Wi
sco
ns
in,
ho
us
i
ng,
m
achine,
tr
ia
zi
nes,
a
uto
are
der
i
ved
f
r
om
[1
8
]
.
The
rest
of
the
da
t
aset
s
are
extra
ct
ed
from
UCI.
Table
1
sho
ws
the
desc
rip
ti
on
of
dataset
.
I
niti
al
l
y,
these
datase
ts
do
no
t
repre
sent
ordi
nal
cl
assifi
cat
ion
,
bu
t
it
rep
rese
nts
r
egr
es
sio
n.
T
o
evo
l
ve
this
re
gr
e
ssio
n
into
ordinal
cl
assifi
cat
ion
we
hav
e
c
on
si
der
e
d
th
e
desire
d
r
esult
is
cat
eg
or
iz
ed
int
o
five
or
te
n
cl
asses
with
e
qual
f
reque
ncy.
Re
gardin
g
the
exp
e
rim
ental
s
et
up
,
a
ho
l
dout
strat
ified
te
c
hniq
ue
was
a
pp
l
ie
d
to
div
ide
the
dataset
s
10
tim
es,
us
in
g
75
pe
rce
nt
of
the
patte
rn
s
f
or
trai
ning
an
d
the
rem
ai
nin
g
25
per
c
ent
for
te
sti
ng
.
Finall
y, the
res
ults
are
taken as t
he
m
ean a
nd stan
da
rd d
e
viati
on of
the m
easur
es
over
the
10 test
set
s.
Table
1.
Natu
r
e of
Dataset
Dataset
Total
n
o
.
o
f
p
atterns
No
.
o
f
Attribu
tes
Total n
o
.
o
f
classes
IR valu
e per c
lass
b
o
n
d
rate
57
37
4
1
.85
,0.1
9
,.
0
.92
,
2
.3
8
Au
to
392
7
5
0
.65
,0.4
0
,0.5
8
,1.1
4
,
7
.15
au
to
m
o
b
ile
205
71
6
1
2
.58
,1.4
3
,0.3
3
,0.4
7
,0.9
0
,1.1
1
Car
1728
21
4
0
.11
,0.8
8
,
5
.98
,6.3
6
ERA1
v
s2
3
4
5
v
s7vs
8
v
s9
1000
4
5
1
.97
,0.0
6
,
2
.07
,6.3
2
,10
.51
Eucalyptu
s1
2
3
v
s4vs
5
736
91
3
0
.25
,0.8
2
,
2
.00
m
a
ch
in
e5
209
6
5
0
.07
,1.3
6
,
2
.92
,
6
.0
4
,
4
.26
m
a
ch
in
e1
0
209
6
10
0
.08
,0.4
6
,0.9
4
,
3
.0
2
,
2
.50
,3.8
0
,7.7
0
,
5
.10
,5.1
0
,3.8
0
triazines5
186
60
5
5
.36
,3.2
7
,1.2
6
,.
0
2
3
,0.4
6
wisco
n
sin
5
194
32
5
0
.38
,0.7
4
,0.7
1
,1.1
4
,
1
.87
wisco
n
sin
1
0
194
32
10
0
.31
,0.8
1
,0.5
9
,1.3
5
,0.7
1
,1.0
2
,1.3
5
,
1
.71
,1.9
7
,1.9
7
h
o
u
sing5
506
13
5
1
.11
,0.2
2
,0.6
2
,
2
.6
1
,3.1
0
Toy
300
2
5
1
.53
,0.4
9
,0.5
6
,0.6
8
,
1
.68
SW
D
1000
4
9
7
.56
,
0
.46
,0.3
8
,0.9
0
The
E
ntire
w
ork
is
validat
ed
base
d
on
A
da
pt
ive
Data
Str
uc
ture
ba
sed
O
ve
rsam
pling
al
gorithm
with
ELM
OP
Cl
assi
fier a
nd these
re
su
lt
s ar
e
co
m
par
e
d Gr
a
ph
ba
sed
oversam
pling
alg
ori
thm
with ELM
OP.
4
.
1.
Per
fo
r
m
an
ce
Me
as
ure
s
This
wor
k
pr
e
ferred
m
os
t
relevan
t
perf
or
m
ance
m
easur
es
su
c
h
as
Me
a
n
Ab
s
ol
ute
Er
ror
(MAE
)
,
Ma
xim
u
m
Mean
Ab
s
olu
te
E
rror
(MM
AE
),
Ge
om
et
ric
M
ean
(
GM)
,
Co
hen’s
Ka
pp
a
a
nd
Acc
ur
a
cy
use
d
to
validat
e
the
pr
opos
e
d wor
k.
4
.
1.
1.
Mean
Ab
s
olut
e
Erro
r
MAE
is
the
a
ve
rag
e
dif
fer
e
nc
e
betwee
n
tr
ue
value
a
nd
e
va
luate
d
value.
MAE
is
the
e
s
sentia
l
an
d
cl
ear m
easur
e
of ave
rag
e
er
ror
[
19]
.
q
N
i
i
i
q
q
y
y
N
M
A
E
1
)
ˆ
(
)
(
1
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ad
ap
ti
ve
Da
t
a Struct
ur
e B
as
e
d Overs
amplin
g
Al
go
rit
hm f
or Or
dinal Cl
assi
fi
cation
(
D.D
hanalaks
hmi
)
1067
4
.
1.2
.
M
ax
im
u
m Me
an A
bsol
ut
e Err
or
Pr
op
os
e
d
MM
AE
m
et
ric
for
ordi
nal
cl
assifi
cat
ion
.
It
disp
la
ys
the
m
axi
m
u
m
MAE
f
or
al
l
the cla
sses
[20
]
.
MM
AE=
Q
q
M
A
E
q
,
.
...
,
1
;
m
a
x
(8)
4
.
1.3
.
Geo
met
ri
c Mean
Geo
m
et
ric
m
e
an
is
on
e
of
t
he
prefe
rab
le
m
easur
es
f
or
i
m
balanced
le
arn
i
ng.
Ge
om
etr
ic
m
ean
i
s
def
i
ned as
fo
ll
ow
s
:
GMean=
FP
TN
TN
FN
TP
TP
`
(9)
4
.
1.4
.
Kapp
a
Coh
e
n’s
ka
pp
a
sta
ti
sti
c
is
on
e
of
the
pref
era
ble
m
easur
es
fo
r
im
balanced
m
ul
ti
cl
ass
le
arn
in
g.
Whe
n
kappa v
al
ue
<
0
is
in
dicat
in
g
no
coe
xists between
act
ual
a
nd p
re
dicte
d
val
ue
,
0
–
0.20
as sli
gh
t
c
oex
ist
s,
0.2
1
–
0.40
as
fair
c
oin
ci
de
,
0.4
1
–
0.60
as
m
od
er
at
e
agr
eem
ent,
0.61
–
0.80
as
su
bs
ta
ntial
,
and
0.8
1
–
1
as
alm
os
t
perfect
agreem
ent.
Kappa=
m
i
i
ci
ri
m
i
i
m
i
i
ci
ri
T
T
N
T
T
TP
N
1
2
1
1
(10)
Wh
e
re
N
total
nu
m
ber
of
pat
te
rn
s,
ri
T
num
ber
of
rows
from
t
he
co
nfusi
on
m
at
rix,
ci
T
num
ber
of
col
um
ns
from
the co
nf
usi
on m
at
rix.
4
.
1.5
.
Accur
ac
y
Accuracy
is th
e pro
portio
n of t
ru
e
r
es
ults, ei
ther
tr
ue p
os
it
ive
or tru
e
n
e
ga
ti
ve.
Accuracy
=(T
P
+TN) /
(
T
N+T
P+FN
+
FP)
(11)
Af
te
r
eval
uatin
g
the
m
easur
es
MAE,
MM
AE
,
GMea
n,
Ka
ppa
a
nd
Acc
ur
a
cy
Ob
ta
ine
d
O
ver
10
R
uns
for
the
E
xisti
ng
Gr
a
ph Base
d O
ver
sam
pling an
d
P
r
opos
e
d ADSO
S r
e
su
lt
s
are display
ed
i
n
Ta
ble
2
-
Ta
bl
e
6.
Table
2
. M
AE M
ean a
nd Stan
dard
De
viati
ons (
Me
a
n±
S
D
)
Dataset/
M
eth
o
d
Graph
Based
Over
sa
m
p
l
in
g
ADSOS
b
o
n
d
rate
0
.17
3
3
±
0
.18
4
2
0
.09
3
3
±
0
.03
2
6
au
to
0
.34
0
9
±
0
.03
5
4
0
.34
0
7
±
0
.03
5
3
au
to
m
o
b
ile
0
.45
7
7
±
0
.27
2
4
0
.39
4
2
±
0
.24
6
9
car
0
.22
5
0
±
0
.02
0
0
0
.22
4
0
±
0
.02
0
1
ERA1
v
s2
3
4
5
v
s7vs
8
v
s9
0
.14
0
4
±
0
.04
4
0
0
.12
7
2
±
0
.04
4
2
Eucalyptu
s1
2
3
v
s4vs
5
0
.30
9
7
±
0
.05
5
6
0
.24
4
5
±
0
.07
3
0
m
a
ch
in
e5
0
.26
4
1
±
0
.04
6
2
0
.27
3
5
±
0
.13
1
3
m
a
ch
in
e1
0
0
.58
4
9
±
0
.17
3
6
0
.52
1
9
±
0
.09
7
8
triazines5
0
.36
8
8
±
0
.02
6
5
0
.36
8
0
±
0
.02
6
9
wisco
n
sin
5
0
.36
0
5
±
0
.03
4
7
0
.44
8
9
±
0
.06
7
0
wisco
n
sin
1
0
1
.13
6
0
±
0
.03
4
7
1
.02
7
1
±
0
.06
8
7
h
o
u
sing5
0
.15
4
8
±
0
.04
2
7
0
.20
7
8
±
0
.08
2
3
Toy
0
.56
8
8
±
0
.03
4
9
0
.54
6
6
±
0
.04
5
8
SW
D
0
.28
0
0
±
0
.00
9
7
0
.20
9
3
±
0
.07
3
2
Table
3
.
MM
A
E Mea
n
a
nd St
and
a
r
d Dev
ia
ti
on
s
(
Me
a
n±SD
)
Dataset/
Metho
d
Graph
Based
Over
sa
m
p
l
in
g
ADSOS
b
o
n
d
rate
1
.05
0
0
±
0
.15
0
0
1
.00
0
0
±
0
.00
0
0
au
to
0
.96
1
3
±
1
.00
2
9
1
.03
5
1
±
0
.02
6
2
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
1
0
63
–
1
0
70
1068
au
to
m
o
b
ile
2
.29
4
1
±
0
.46
2
4
2
.30
0
0
±
0
.45
8
2
car
0
.32
0
7
±
0
.02
8
6
0
.30
4
2
±
0
.02
8
6
Dataset/
Metho
d
Graph
Based
Over
sa
m
p
l
in
g
ADSOS
ERA1
v
s2
3
4
5
v
s7vs
8
v
s9
1
.10
4
3
±
0
.11
6
9
1
.07
3
8
±
0
.09
5
3
Eucalyptu
s1
2
3
v
s4vs
5
0
.54
6
7
±
0
.10
0
0
0
.43
1
0
±
0
.12
8
1
m
a
ch
in
e5
0
.36
8
4
±
0
.06
4
4
0
.38
1
5
±
0
.18
3
1
m
a
ch
in
e1
0
1
.06
8
9
±
0
.31
7
2
0
.95
3
9
±
0
.17
8
8
triazines5
3
.00
0
0
±
0
.00
0
0
3
.00
0
0
±
0
.00
0
0
wisco
n
sin
5
1
.04
2
9
±
0
.10
5
0
1
.16
5
6
±
0
.09
7
9
wisco
n
sin
1
0
3
.11
6
1
±
0
.03
5
7
2
.77
7
0
±
0
.17
4
5
h
o
u
sing5
0
.61
4
0
±
0
.31
0
8
0
.85
2
6
±
0
.23
7
5
Toy
2
.15
7
4
±
0
.11
1
8
2
.05
9
5
±
0
.22
8
4
SW
D
1
.4
5
8
3
±
0
.1
5
5
9
1
.5
8
3
3
±
0
.1
3
8
1
Table
4
. GM
Me
an
a
nd Stan
dard
De
viati
ons (
Me
a
n±
S
D
)
Dataset/
Metho
d
Graph
Based
Over
sa
m
p
l
in
g
ADSOS
b
o
n
d
rate
0
.84
0
3
±
0
.07
3
3
0
.86
7
9
±
0
.01
8
6
au
to
0
.83
3
1
±
0
.12
2
4
0
.83
3
1
±
0
.00
9
2
au
to
m
o
b
ile
0
.74
1
9
±
0
.04
7
6
0
.74
3
3
±
0
.04
8
6
car
0
.96
5
2
±
0
.00
3
0
0
.96
5
2
±
0
.00
3
0
ERA1
v
s2
3
4
5
v
s7vs
8
v
s9
0
.84
0
9
±
0
.02
9
0
0
.85
3
6
±
0
.00
5
1
Eucalyptu
s1
2
3
v
s4vs
5
0
.83
5
7
±
0
.03
0
0
0
.87
0
5
±
0
.03
8
5
m
a
ch
in
e5
0
.96
5
7
±
0
.00
8
1
0
.96
0
9
±
0
.01
7
9
m
a
ch
in
e1
0
0
.97
5
3
±
0
.00
6
6
0
.97
3
2
±
0
.01
6
7
triazines5
0
.60
4
3
±
0
.00
1
2
0
.60
2
9
±
0
.00
2
9
wisco
n
sin
5
0
.85
3
5
±
0
.00
7
9
0
.82
4
5
±
0
.02
2
5
wisco
n
sin
1
0
0
.75
9
4
±
0
.07
6
4
0
.81
6
4
±
0
.03
1
1
h
o
u
sing5
0
.90
2
3
±
0
.03
5
4
7
0
.86
4
3
±
0
.04
1
9
Toy
0
.72
5
8
±
0
.00
3
8
0
.72
8
9
±
0
.00
9
1
SW
D
0
.7
2
2
7
±
0
.0
0
8
4
0
.7
7
0
8
±
0
.0
4
9
2
Table
5
. Ka
pp
a
Mea
n
a
nd Sta
nd
a
r
d Dev
ia
ti
ons
(Mean
±S
D)
Dataset/
Metho
d
Graph
Based
Ov
er
sa
m
p
l
in
g
ADSOS
b
o
n
d
rate
0
.64
1
0
±
0
.21
4
2
0
.72
0
5
±
0
.08
9
2
au
to
0
.08
1
9
±
0
.04
8
9
0
.08
1
5
±
0
.04
8
8
au
to
m
o
b
ile
0
.37
8
4
±
0
.03
8
7
0
.36
7
0
±
0
.04
0
9
car
0
.70
0
0
±
0
.02
6
7
0
.70
0
0
±
0
.02
6
7
ERA1
v
s2
3
4
5
v
s7vs
8
v
s9
0
.59
1
2
±
0
.10
6
3
0
.64
3
7
±
0
.09
0
8
Eucalyptu
s1
2
3
v
s4vs
5
0
.30
3
0
±
0
.12
5
1
0
.44
9
7
±
0
.16
4
3
m
a
ch
in
e5
0
.56
7
6
±
0
.10
0
2
0
.51
0
3
±
0
.22
5
3
m
a
ch
in
e1
0
0
.21
0
9
±
0
.12
8
4
0
.43
9
0
±
0
.04
8
0
triazines5
0
.24
6
4
±
0
.03
1
3
0
.25
5
3
±
0
.02
6
6
wisco
n
sin
5
0
.05
8
4
±
0
.02
7
1
0
.17
6
2
±
0
.07
6
6
wisco
n
sin
1
0
0
.62
1
8
±
0
.00
7
5
0
.60
2
6
±
0
.03
4
2
h
o
u
sing5
0
.54
0
6
±
0
.11
4
2
0
.42
4
8
±
0
.14
7
7
Toy
0
.22
5
2
±
0
.02
0
4
0
.21
6
7
±
0
.03
5
1
SW
D
0
.2
9
2
4
±
0
.0
3
6
2
0
.4
9
1
5
±
0
.2
0
5
1
Table
6
. Acc
uracy
Mea
n
a
nd
Stand
a
r
d Dev
i
at
ion
s
(Mean
±
SD
)
Dataset/
Metho
d
Graph
Based
Over
sa
m
p
l
in
g
ADSOS
b
o
n
d
rate
0
.83
3
3
±
0
.16
9
3
0
.90
6
6
±
0
.03
2
6
au
to
0
.70
1
0
±
0
.02
2
3
0
.70
3
0
±
0
.02
3
4
au
to
m
o
b
ile
0
.67
5
0
±
0
.14
4
5
0
.70
3
8
±
0
.13
9
2
car
0
.88
7
5
±
0
.01
0
0
0
.88
7
5
±
0
.01
0
0
E
RA
1
v
s
2
3
4
5
v
s
7
v
s
8
v
s
9
0
.8
6
9
2
±
0
.0
3
4
0
0
.8
8
6
0
±
0
.0
2
9
0
Eucalyptu
s1
2
3
v
s4vs
5
0
.69
0
2
±
0
.05
5
6
0
.87
0
5
±
0
.07
3
0
m
a
ch
in
e5
0
.86
1
6
±
0
.03
2
0
0
.84
3
4
±
0
.07
2
1
m
a
ch
in
e1
0
0
.83
0
2
±
0
.04
6
2
0
.81
7
6
±
0
.11
1
4
triazines5
0
.75
8
9
±
0
.01
0
0
0
.76
1
7
±
0
.00
8
5
wisco
n
sin
5
0
.65
9
9
±
0
.00
9
6
0
.60
8
1
±
0
.03
6
3
wisco
n
sin
1
0
0
.52
3
8
±
0
.00
9
6
0
.54
4
2
±
0
.03
7
2
h
o
u
sing5
0
.85
3
0
±
0
.03
6
5
0
.79
9
9
±
0
.08
1
3
Toy
0
.58
6
6
±
0
.01
0
9
0
.59
0
6
±
0
.01
7
9
SW
D
0
.7
3
4
6
±
0
.0
1
3
5
0
.8
0
9
3
±
0
.0
7
6
9
To
qua
ntify
wh
et
her
a
sta
ti
sti
cal
diff
ere
nce
exists
am
ong
the
al
go
r
it
h
m
s
co
m
pared,
t
-
Test
is
perform
ed
on t
he
m
ean r
a
nk
i
ng of all
the
eva
luati
on
m
easu
res
it
is
disp
la
ye
d
in
Ta
ble 7.
Evaluation Warning : The document was created with Spire.PDF for Python.
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci
IS
S
N:
25
02
-
4752
Ad
ap
ti
ve
Da
t
a Struct
ur
e B
as
e
d Overs
amplin
g
Al
go
rit
hm f
or Or
dinal Cl
assi
fi
cation
(
D.D
hanalaks
hmi
)
1069
Table
7
. t
-
Test
on Mea
n R
an
ki
ng
of the
Eval
uation M
easu
r
es (
α
=0
.05
)
Ob
serv
atio
n
s
Graph
Based
Over
sa
m
p
l
in
g
ADSOS
Mean
1
.53
8
1
.23
8
Variance
0
.03
7
2
7
0
.00
6
3
7
Ob
serv
atio
n
s
5
5
Pearso
n
Co
rr
elatio
n
-
0
.66
5
5
5
8
5
2
7
Hy
p
o
th
esized
M
ea
n
Dif
f
erence
0
df
4
t Stat
2
.64
8
5
4
8
4
7
8
P(T
<=t
)
o
n
e
-
tail
0
.02
8
5
3
4
1
1
5
t Critical
on
e
-
tail
2
.13
1
8
4
6
7
8
6
P(T
<=t
)
two
-
tail
0
.05
7
0
6
8
2
3
1
t Critical
two
-
tail
2
.77
6
4
4
5
1
0
5
The
te
st
pro
ve
s
that,
null
hy
po
the
sis
is
r
ejected
w
her
e
p
val
ue
is
le
ss
than
(α=0
.
05)
that
tw
o
al
gorithm
s
perform
s
si
m
il
arl
y
in
m
ean
rank
ing
of
the
eval
uation
m
easur
e
s
how
eve
r
ADSOS
pe
rfor
m
s
bette
r
than
G
raph Ba
sed Ove
rsam
pling
Met
hod wi
th ELM
OP
as
cl
assifi
er whic
h
is
de
picte
d
in
Fig
ur
e
1.
Figure
1.
Me
an
ra
nk
i
ng of the
evaluati
on m
easur
es
5.
CONCLUS
I
ON
In
this
pa
per
we
pro
po
se
d
the
novel
A
da
ptive
Data
St
r
uctu
re
base
d
Ov
e
rsam
pling
al
gorithm
to
pr
e
fer
t
he
use
fu
l
obje
ct
s
for
fu
rt
her
pr
oce
ssing.
W
e
co
m
par
ed
our
m
et
hods
with
e
xisti
ng
gra
ph
base
d
pr
e
processi
ng
al
gorithm
fo
r
four
te
e
n
data
set
s.
O
ur
ai
m
of
t
his
w
ork
i
s
to
com
par
e
t
he
pr
opos
e
d
A
DSOS
pr
e
processi
ng
al
gorithm
wit
h
e
xisti
ng
pre
processi
ng
al
gorithm
for
or
din
al
cl
assifi
c
at
ion
.
W
it
h
r
egards,
we
ad
op
t
a
ny
on
e
of
the
ordi
nal
cl
assifi
er
s
uch
as
ELM
O
P
to
validat
e
our
work.
Th
us
our
pro
pose
d
m
et
ho
d
on
ly
oversam
ples
obj
ect
s
wh
i
ch
ha
ve
highes
t
con
fi
den
ce
a
nd
c
om
plica
ted
re
gions.
E
xperim
ents
ind
ic
at
e
that
ou
r
m
et
ho
d be
hav
e
s
bette
r
in
te
rm
s o
f
er
ror r
at
e, accu
racy s
ensiti
vity
.
ACKN
OWLE
DGE
MENTS
I
w
ou
l
d
li
ke
to
ex
pr
es
s
m
y
sp
eci
al
gr
at
it
ud
e
an
d
tha
nk
s
t
o
D
r.K.
Karu
nak
a
ran
,
Pr
inci
pal
an
d
Secretary
,
Sr
i
Ram
akr
ish
na
Coll
ege
of
Ar
t
s
an
d
Scie
nce,
Coim
bator
e
f
or
pro
vid
i
ng
e
xcell
ent
in
fr
ast
ru
ct
ur
e
and
s
upport
for
m
y
Re
search
w
ork.
I
am
hig
hly
i
nd
e
bted
to
m
y
Re
search
Gu
i
de
Dr.A
nn
a
Sa
ro
Vi
j
e
ndra
n,
Dean
Sc
hool
of
Com
pu
ti
ng,
Sr
i
Ram
a
kr
ish
na
Coll
ege
of
Ar
ts
an
d
Scie
nc
e,
Coim
bator
e
fo
r
t
heir
gui
dan
ce
,
const
ant
super
visio
n,
sup
port,
effor
t,
in
valu
ably
con
str
uctive
crit
ic
is
m
an
d
fr
ie
nd
ly
ad
vice
fo
r
m
y
resear
c
h
work.
REFERE
NCE
S
[1]
Qiong,
G.
,
Ca
i,
Z.
,
Zhu
,
L.
and
Huang,
B
.
,
Dat
a
Mini
ng
on
Im
balanc
ed
Data
Set
s
,
In
te
rn
at
ion
al
Conf
ere
n
ce
o
n
Advanc
ed
Com
p
ute
r Theor
y
and
Engi
ne
eri
ng,
Ph
uket
,
20
-
22
De
c
ember,
2008
;102
0
-
1024.
[2]
Zhuo
y
u
an
Zh
en
g
,
Yunpeng
C
a
i,
Ye
Li,
Over
Sam
pli
ng
Metho
d
For
Im
bal
ance
d
Cla
ss
ifi
c
at
ion
,
Computing
and
Informatic
s
,
201
5;
34
:1017
–
1037
.
[3]
Rohini
R.
Rao
,
Krishnam
oorthi
Makkit
ha
y
a
,
Learni
ng
from
a
C
la
ss
Im
bal
anced
Public
H
ea
l
th
Data
set
:
a
Cost
-
base
d
Com
par
ison
of
Cla
ss
ifi
er
Perform
anc
e,
I
nte
rnational
Jo
urnal
of
Elec
tri
cal
and
Compu
te
r
Engi
n
ee
rin
g
(
IJE
CE)
,
2017;7
(4):
.
2215
-
2222
.
0
0
.5
1
1
.5
2
Gr
ap
h
b
ased
Ov
er
s
a
m
p
lin
g
AD
SOS
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2502
-
4752
Ind
on
esi
a
n
J
E
le
c Eng &
Co
m
p
Sci,
Vo
l.
12
, N
o.
3
,
Dece
m
ber
2
01
8
:
1
0
63
–
1
0
70
1070
[4]
X
ia
oQing
Gu,
TongGuang
Ni,
W
ei
W
ang,
Online
Im
balanc
e
d
Support
Vec
t
or
Mac
hine
for
Phis
hing
Emai
ls
Filt
eri
ng
,
TEL
K
OMNIKA
Indon
esian
Journal
of
El
e
ct
rica
l
Eng
in
ee
ring,
2014;12(
6):
4306
–
4313.
[5]
Xiaoz
hi
Du,
Huim
in
Lu,
Gan
g
Li
u,
Softwar
e
Aging
Predi
ct
ion
b
ase
d
on
Ext
reme
L
ea
r
ning
Mac
hin
e
,
TEL
KOMNIKA,
2013;11(11):
65
47
-
6555.
[6]
Mohd
Abdul
Tal
ib
Mat
Yus
oh,
Saidatul
Habsah
As
m
an,
Zuha
ila
Mat
Yasin,
Ahm
ad
Farid
Abidin,
Cla
ss
ifi
c
ation
o
f
The
NT
EV
Problems
on
the
Co
m
m
erc
ia
l
Bui
lding,
Indone
sian
Journal
of
Elec
t
rical
Eng
i
ne
erin
g
and
Computer
Sci
en
ce,
2018;9(
2):
380
-
386.
[7]
Marıa
Per
ez
-
Ort
iz
,
Pedro
Anton
io
Gutie
rr
ez,Ce
s
ar
Herva
s
-
Mart
ı
nez
and
Xin
Ya
o,
Graph
B
ase
d
Approac
hes
for
Over
-
sam
pli
ng
in
the
Conte
xt
o
f
Ordina
l
Regres
sion,
IEE
E
Tr
ansacti
ons
on
Knowle
dge
and
Data
Engi
ne
erin
g
,
2015;
27(5):123
3
-
1245.
[8]
Anna
Saro
Vije
ndra
n,
D.
Dhan
al
akshm
i,
An
a
naly
s
is
of
Data
Chara
c
te
rist
ic
s
a
nd
Cla
ss
ifi
er
Pe
rform
anc
e
in
th
e
Conte
xt
of
Ordina
l
Cla
ss
ifi
c
atio
n,
Inte
rnationa
l
journal
of
Data
mining
and
Eme
rging
Technol
ogie
s
,
2016;
7(1)
:
36
-
42.
[9]
Ann
a
Saro
Vije
n
dra
n,
D.
Dhan
alakshm
i,
A
Novel
Approac
h
in
O
ver
sam
pli
ng
Alg
orit
hm
for
Im
bal
anc
ed
Da
ta
Sets
in
the
context
of
Ordina
l
Cla
s
sific
a
ti
on,
I
EE
E
int
ernati
onal
Confe
renc
e
on
Computati
onal
Inte
lligen
ce
and
Computing
R
ese
arch
,
Chenn
ai,2
016;
964
-
968.
[10]
Anna
Saro
Vijendran,
D.
Dhan
al
akshm
i
,
An
Im
prove
d
Sam
pl
ing
Algorit
hm
for
Im
bal
anc
ed
Data
Sets
in
th
e
cont
ex
t
of
Ordina
l
Cla
ss
ifica
t
io
n,
Inte
rn
at
ion
al
journa
l
of
Adva
nce
Resea
r
ch
in
Scie
nce
and
E
ngine
er
ing,
201
7;
6(12):201
-
209.
[11]
Im
an
Nekooe
imehr,
Sus
ana
K.
La
i
-
Yuen
,
Clust
er
-
base
d
W
ei
ght
ed
Oversam
pli
ng
for
Ordina
l
Regre
ss
ion
(CW
O
S
-
Ord),
Journal
N
euroc
omputing
,
2016;
218(C):51
-
60.
[12]
H.
He,
Y.
Bai,
E.
Gar
ci
a
,
S.
Li,
AD
ASYN,
Adapt
ive
synth
et
i
c
sam
pli
ng
approach
for
imbalan
c
ed
le
arning
”,
in:
Proce
edi
ngs
of
t
he
IEEE
In
t
ern
a
ti
onal
Joint
Con
fer
ence
on
Neur
al
Networks,
IJCN
N,
IEE
E
W
o
rld
Congress
on
Com
puta
ti
onal I
nte
lligen
ce, 200
8;1322
-
1328.
[13]
Im
an
Nekooe
imehr,
Sus
ana
K
,
Adapti
ve
sem
i
-
u
nsupervise
d
wei
ghte
d
over
sam
pl
ing
(A
-
SU
W
O)
for
imbala
n
ce
d
dat
ase
ts,
Ex
pert
Syste
ms
wit
h
A
p
pli
cations
,
2016
;
46:405
-
416.
[14]
Jos
e
A.Sae
z
,
Ba
rtosz
Krawczy
k
,
Micha
l
W
ozni
a
k,
Anal
y
sisng
th
e
over
sam
pli
ng
of
diffe
r
ent
class
es
and
t
y
p
es
of
exa
m
ple
s in
m
ul
ti
-
class
imbala
n
c
ed
da
ta
set
,
Patte
rn R
ec
ogn
it
ion
,
2016;
57:164
-
17
8.
[15]
H.
Han,
W
.
Y.
W
ang,
and
B.
H.
Mao,
Borderli
n
eSM
OT
E:
A
Ne
w
Ove
r
-
Samplin
g
Me
thod
in
Im
balanc
ed
Data
S
et
s
Learning
,
In
Proce
ed
ings
of
th
e
Inte
rn
at
ion
al
Confer
ence
on
I
nte
lligent
Com
puti
ng
2005;
Par
t
I,
LNCS
3644
,
878
–
887.
[16]
F.
Ferna
nde
z
-
Na
var
ro,
C
.
Herv
as
-
Martı
ne
z,
and
P.
A.
Guti
err
e
z,
A
D
y
nami
c
Ove
r
-
Sam
pli
ng
Proc
edur
e
B
ase
d
on
Sensiti
vity
for M
ult
i
-
class
Problems
,
Pattern
Recogni
t
.
,
2011;
44
(8):1821
–
1833.
[17]
Ferná
ndez
A.
,
del
Jesus
M.J.,
Herre
ra
F.,
Mul
ti
-
cl
ass
Imbalance
d
Data
-
Set
s
wit
h
Linguisti
c
Fuzzy
Rul
e
Based
Classifi
cation
Syste
ms
Based
on
Pai
rwis
e
Learning
,
In:
Hülle
rm
eier
E.,
Krus
e
R.
,
Hoffm
ann
F.
(ed
s)
Com
puta
ti
onal
I
nte
lligen
ce
for
Know
le
dge
-
Bas
ed
S
y
stems
Design.
IPM
U,
Le
c
ture
Notes
in
C
om
pute
r
Scie
nc
e
,
Springer,
B
erlin,
Heidelbe
rg
201
0;
6178.
[18]
W
.
Chu
and
Z
.
Ghahra
m
ani
,
Ga
uss
ia
n
proc
esses
for
ordin
al re
gr
e
ss
ion,
J. M
ach
.
L
earn.
R
es.
,
2005;
6:1019
–
1041.
[19]
Cort
J.
W
i
ll
m
ott,
Kenj
i
Ma
tsuura,
Adv
ant
ag
es
o
f
the
m
ea
n
absol
ute
err
or
(MA
E)
over
the
root
m
ea
n
squar
e
err
or
(RMS
E)
in
asses
sing a
ver
age model
p
erf
orm
anc
e
,
Cli
mat
e
R
ese
ar
ch
,
2005;
30:
79
–
82.
[20]
M
.
Cruz
-
Ramırez
,
C
.
Herva
s
-
Martı
ne
z,
J.
Sa
nche
z
-
Moned
ero
,
and
P.
A.Guti
err
ez,
Metrics
t
o
guide
a
m
ult
i
-
obje
c
ti
ve
evo
lut
i
onar
y
a
lgori
thm
for
ordinal
cl
assi
ficat
ion
,
Neuroco
mputing
,
2014
;
1
35:21
–
31.
Evaluation Warning : The document was created with Spire.PDF for Python.