Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
9
, No
.
5
,
Octo
ber
201
9,
pp. 364
9
~3
656
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v9
i
5
.
pp3649
-
36
56
3649
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Estima
ti
on of
regression
-
based
m
odel
wit
h
bu
lk
no
is
y dat
a
Chanint
orn Ji
ttawi
ri
ya
n
uk
oon
Gradua
te School
of
Advan
ce
d
Technol
og
y
Man
ag
ement
,
As
sum
pti
on
Univer
sit
y
,
T
hai
l
and
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Sep
18
, 201
8
Re
vised
A
pr
8
,
201
9
Accepte
d
Apr
17
, 201
9
The
bu
lk
nois
e
has
b
ee
n
p
rovoking
a
co
ntri
bute
d
da
ta
due
to
a
comm
unic
at
ion
net
work
with
a
tre
m
endo
usl
y
low
signal
to
noise
rat
io
.
An
appr
ec
i
at
ed
m
et
hod
for
rev
i
sing
m
assive
no
ise
of
indi
viduals
through
informati
on
th
eo
r
y
is
widely
dis
cussed.
One
of
t
he
pra
ct
i
ca
l
app
l
ic
a
ti
ons
of
thi
s
appr
oac
h
for
bulk
noise
esti
m
at
ion
is
ana
l
y
z
ed
using
int
el
l
ige
n
t
aut
om
at
ion
and
m
ac
hine
learni
ng
tool
s,
dealin
g
the
ca
se
of
bul
k
noise
exi
sten
ce
or
no
nexi
sten
ce
.
A
r
egr
ession
-
base
d
m
odel
is
empl
o
y
ed
for
th
e
inve
stigation
an
d
expe
r
iment.
E
stim
at
ion
for
th
e
pra
ct
i
cal
ca
se
with
bu
lk
nois
y
d
ataset
s
i
s
proposed.
Th
e
proposed
m
ethod
appl
i
es
sli
ce
-
and
-
d
i
c
e
te
chn
ique
to
par
t
it
ion
a
bod
y
o
f
d
at
ase
ts
down
int
o
slight
er
porti
o
ns
so
tha
t
it
ca
n
be
ca
r
rie
d
o
ut.
The
av
era
g
e
err
or,
cor
r
el
a
ti
o
n,
absolut
e
err
o
r
and
m
ea
n
square
err
or
ar
e
computed
to
va
li
date
the
esti
m
a
ti
on.
Resul
ts
from
m
assive
onli
ne
anal
y
sis
will
be
ver
if
ie
d
with
data
co
llected
in
the
fo
ll
owi
ng
per
iod
.
I
n
m
an
y
ca
ses,
the
pre
di
ct
ion
with
bulk
nois
y
d
ata
through
MO
A
sim
ula
ti
on
rev
eals Ra
ndom
Im
puta
ti
on
m
inim
iz
es
the a
v
era
g
e
err
o
r.
Ke
yw
or
d
s
:
Bulk
noise
Cl
assifi
cat
ion
Estim
at
ion
Me
an
s
qu
a
re e
rror
No
isy
a
nd m
is
sing data
Re
gr
essi
on
-
bas
ed
m
od
el
Copyright
©
201
9
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
:
Chani
nto
r
n Ji
tt
awiriy
an
ukoon
,
Gr
a
duat
e Sc
hool of
A
dv
a
nce
d
Tec
hnol
og
y
Ma
nag
em
ent,
Assum
ption
Unive
rsity
,
Sam
ut Pr
aka
n
Pr
ovi
nce, 1
0540, T
haila
nd.
Em
a
il
: pct25
26
@yah
oo.co
m
1.
INTROD
U
CTION
Currentl
y,
noisy
and
m
issi
n
g
dataset
s
are
on
the
rise
due
to
m
a
lfun
c
ti
on
in
sen
sor
te
chnolo
gy.
These
no
isy
pa
rts
distu
rb
par
t
ic
ular
fiel
ds
in
the
ta
ble
datas
et
.
Hu
m
an
being
ca
n
fi
gure
out
what
these
bl
anks
sh
oul
d
sta
nd
for,
or guess
ar
ound them
. Bu
t
m
os
t pr
ogram
m
ers
think
thes
e the h
ar
d
pro
bl
e
m
s ex
act
ly b
ecaus
e
they
ha
ve
no
cl
ue
a
bout
t
he
m
et
hod
to
a
pp
ly
.
The
th
reat
of
f
ixing
noisy
data
is
not
on
ly
a
m
ental
fear
but
al
so
tur
ns
a
sim
ple
com
pu
ta
ti
on
t
o
a
n
e
xtra
cal
c
ulati
on
,
just
only
to
fin
d
ou
t
w
her
e
the
m
i
ssing
data
is
locat
ed
.
Keep
i
ng
this
obser
vatio
n
in
m
ind
,
s
om
e
app
r
oac
hes
wh
ic
h
will
help
cal
cula
ti
on
f
or
da
ta
analy
ti
cs
w
ork
ou
t
wh
il
e s
olv
i
ng
m
issi
ng
d
at
a
pr
ob
le
m
s
m
us
t b
e pro
vid
e
d
s
o
t
hat d
at
a i
nter
pret
at
ion
ca
n be
perform
ed.
Ba
sh
a
et
al
.
[
1]
propose
a
si
m
il
arity
-
based
cl
us
te
r
for
cl
as
sifyi
ng
t
he
stri
ng.
It
cal
culat
es
the
c
os
ine
si
m
il
arity
then
com
par
es
wit
h
the
vect
or
of
the
i
nfor
m
at
i
on
of
the
te
xt.
This
te
c
hn
i
que
is
ap
plied
f
or
te
xt
-
base
d
dataset
s,
an
d
t
he
perform
ance
is
inv
e
sti
gated
usi
ng
sig
nal
to
no
ise
r
at
io,
m
ean
squar
e
erro
r
,
and
exec
utio
n
tim
e.
Lee
et
al
.
[2
]
pro
po
s
e
a
f
uzzy
-
ba
se
d
a
ppro
ac
h
t
o
cl
assify
the
s
entence
te
xt
in
m
any
do
c
um
ent
dataset
s.
A
fu
zzy
m
et
ric
has
be
en
analy
ze
d
to
tra
ns
la
te
the
high
le
vel
in
to
a
low
-
le
vel
te
xt.
The
pro
po
se
d
cl
us
te
r
div
ide
s
the
sta
te
sp
ace
into
s
ub
-
s
pace
s.
T
he
disc
rete
su
b
-
s
paces
a
re
thu
s
unit
ed
to
create
the
se
par
at
e
le
vel.
T
he
a
ppr
oa
ch
pro
du
ces
op
ti
m
iz
ation
a
nd
r
uns
faster
than
oth
e
r
te
c
hniq
ues.
W
ei
et
al
.
[
3]
introd
uce
a
le
xical
-
base
d
se
ntence.
It
em
plo
ys
the
on
t
ology
orde
red
patte
rn
s
t
o
c
om
pu
te
the
sim
il
arity
betwee
n
i
nd
i
vid
ual
w
ords.
T
he
le
xical
-
based
m
et
h
od
is
us
e
d
t
o
a
naly
ze
th
e
sem
antic
of
t
he
wor
ds
.
Com
pare
d
to
the
te
xt
cl
ust
ering,
the
cl
ust
ering
of
the
e
m
pt
y
(b
la
nk)
t
exts
are
t
he
m
os
t
di
ff
ic
ult.
A
s
the
m
issi
ng
words
i
n
the
dataset
a
r
e
widely
e
nc
ounte
red
by
a
ppli
cat
ion
s
i
n
r
eal
it
y,
the
cl
ust
e
ring
of
t
he
noisy
dataset
nee
d
s
the att
entio
n.
I
n
this
p
a
per, th
e pro
posed
m
eth
od
base
d o
n
e
stim
ation
ce
nters
the
case
of
m
issi
ng
d
at
a.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19 :
3649
-
3656
3650
In
real
li
fe
,
noisy
data
[4
]
de
ve
lop
s
wh
e
ne
ve
r
blan
k
has
re
pl
aced
any
entri
es
in
the
ta
ble
as
sh
ow
n
in
F
igure
1.
No
is
e
data
can
de
v
el
op
an
d
root
s
a
n
unreli
abl
e
con
se
quence
ran
gi
ng
from
the
or
igin
al
dataset
cannot
be
e
xec
uted
(err
or
f
ound)
t
o
nonr
e
spon
si
ve
process
w
hile
r
unning
the
fil
e
.
Liu
e
t
al
.
[
5]
intr
oduce
a
tree
-
ba
sed
cl
ust
ering
al
gorit
hm
fo
r
cl
assify
ing
te
xts
in
paral
le
l.
The
p
ara
ll
el
al
go
rithm
per
se
m
ini
m
izes
the
tim
e
con
voluti
on
by
e
xec
utin
g
the
data
c
ollec
ti
on
,
t
he
data
par
ti
ti
onin
g,
and
cl
assifi
cat
ion.
T
he
m
ast
er
-
sla
ve
process
in
paral
le
l
pr
ocessin
g
is
app
li
ed
.
The
pro
pose
d
m
et
ho
d
im
pr
ove
s
tim
e
co
m
plexity
and
ac
cur
acy
.
Sh
i
et
al
.
[6
]
su
ggest
a
ge
ne
ti
c
al
go
rithm
t
o
com
bin
e
the
fitness
eq
uation
to
the
c
onverge
nce
funct
ion
in
K
-
Me
a
ns
cl
us
t
erin
g.
D
oc
ume
nts
base
d
up
on
to
pics
are
su
m
m
arized
t
o
pro
vid
e
a
be
tt
e
r
com
pr
eh
ensio
n.
Li
et
al
.
[
7]
pro
pose
a
F
uzz
y
-
base
d
al
gorithm
to
increase
the
e
ff
ic
ie
ncy
an
d
acc
ur
a
cy
of
the
te
xt
da
ta
set
.
It
pro
ves
that
the
cl
assifi
cat
io
n
is
m
or
e
eff
ec
ti
ve
than
the
re
gu
la
r
te
xt
pa
rtit
ion
in
g
al
gorit
hm
s.
Ba
no
et
al
.
[8
]
dev
el
op
th
e
trai
nin
g
al
go
rith
m
based
upon
the
neural
net
work
m
od
el
.
Ra
nk
i
ng
-
based
c
lusterin
g
an
d
f
il
te
ring
te
chn
iq
ue a
re
use
d
t
o
el
im
inate
the
disti
nct dat
a.
No
isy
data
re
pr
ese
nts
a
w
ort
hless
data.
An
y
il
le
gib
le
data
wh
ic
h
i
s
colle
ct
ed
a
ut
om
a
ti
cal
l
y,
will
ou
tc
om
e
a
nd
ca
n
be
desc
ribe
d
as
no
ise
.
Sh
abi
r
et
al
.
[9
]
accor
d
a
de
no
ise
process
t
o
rec
ov
e
r
the
qual
it
y
of
sat
el
li
te
i
m
a
ge.
Th
ese
n
ois
es
are
i
nclusi
ve
of n
ot only
in
com
patible
pro
blem
s
su
ch
as
h
ar
dware or
s
oft
war
e
incom
patibil
it
y
bu
t
al
s
o
pr
oce
ssing
haza
rd
s
su
c
h
as
no
e
xe
cution,
no
ope
rati
on
,
re
j
ect
io
n
or
fail
ur
e.
A
li
tt
l
e
portio
n
of
no
is
e
can
sim
ply
s
cram
ble
the
cl
us
te
rin
g
proce
ss
of
m
easur
ed
data.
It
is
m
or
e
tro
uble
in
c
ase
of
i
m
age
data
from
sat
ellit
e
as
it
connect
s
with
geost
at
ion
ary.
If
an
al
gorithm
fo
cuses
on
a
di
gital
trans
form
ation
,
it
qu
antifi
es
noise
into
erro
r
data.
If
it
is
t
o
be
serio
us
ly
ha
rsh
to
no
ise
,
it
has
to
de
no
ise
then
fine
-
t
un
e
. Ho
w
ever, in
the ca
s
e of
bu
l
k no
isy
d
at
a, it
worse
ns t
he ori
gin
al
it
y and cau
ses a
gr
eat
haza
rd.
F
i
gu
r
e
1.
E
xa
m
pl
e
of
no
i
s
e
da
t
a
(
bl
a
nk
)
The
ob
j
ect
ive
of
this
pa
per
is
to
e
stim
a
te
th
e
accuracy
of
t
he
re
gr
essi
on
-
base
d
m
od
el
for
bulk
no
is
e
data
us
i
ng
MO
A
[
10
]
sim
ulatio
n.
I
n
m
any
cases
of
bu
l
k
no
ise
data,
no
isy
par
t
is
s
uc
h
a
l
arg
e
fr
act
i
on
(
above
sixty
per
ce
nt)
of
t
he
total
dat
a
siz
e.
The
rea
so
n
w
hy
it
'
s
c
al
le
d
"
bu
l
k
noi
se"
is
tha
t
irra
ti
on
al
fluct
uation
is
com
m
on
ly
du
e
to
va
riables
wh
ic
h
are
una
ble
to
be
acco
un
te
d
f
or.
Bul
k
noise
will
be
serio
us
ly
ta
ke
n
into
account
f
r
om
m
any
pr
act
ic
al
cases
.
Firstl
y,
the
noisy
det
erm
inistic
par
t
will
be
in
quir
ed
t
o
br
ea
k
t
hro
ugh
the
error
in
m
a
ni
pula
ti
on
.
Sec
ondly,
the
pro
pose
d
est
i
m
ation
will
treat
these
blan
k
data
then
re
gressi
on
-
base
d
resu
lt
s
f
ro
m
si
m
ulati
on
are
us
e
d
to
valida
t
e
the
pr
op
os
e
d
m
et
ho
d.
La
stl
y,
t
he
pr
eci
sion
of
bu
l
k
no
is
e
treatm
ent
will
b
e
discusse
d af
te
r
com
par
i
ng
with the
actual
dataset
.
2.
RE
LATE
D
W
ORK
The
no
isy
data
is
exactl
y
fu
z
zy
as
it
co
m
pr
ise
s
of
a
ny
val
ue
(i
nclu
ding
blan
k)
wh
ic
h
i
s
unreli
able
and
un
us
ed
.
It
co
uld
be
e
qu
i
pm
ent
err
ors,
hu
m
an
er
rors,
and
im
pr
op
e
r
data
ent
ry.
N
oi
sy
data
is
co
m
m
on
ly
fou
nd
as
m
edical
dataset
s
are
held
.
T
hese
va
lues
dev
el
op
unkn
own
c
halle
ng
e
s
to
the
dat
a
sci
entist
.
A
m
et
hod
for
treat
ing
the
m
issi
ng
value
s
is
pr
opose
d
[
11
]
as
the
chara
ct
erist
ic
values
are
natural
ly
recurri
ng
us
in
g
non
-
par
am
et
ric
discreti
zat
ion
ap
proac
h.
T
he
c
oncept
of
z
-
sc
ore
to
tre
at
the
se
m
issi
ng
pa
r
ts
in
the
datas
et
s
is
app
li
ed
.
Th
e
r
equ
i
rem
ent
is
on
ly
that
data
need
s
t
o
be
r
ecurred
.
The
a
uthors
[12]
pr
opos
e
te
ch
nique
of
handlin
g
no
ise
data
occurre
d
in
m
edical
dataset
s
by
cl
assify
ing,
est
im
a
tin
g,
a
nd
cl
us
te
r
ing
.
O
ne
of
th
e
vita
l
t
echn
iq
ues
in
sta
ti
sti
cs
is
to
fin
d
the
locat
i
on
of
al
l
deter
m
inist
ic
par
ts
wh
i
c
h
will
be
us
e
d
f
or
est
im
at
ion
.
In
m
any
cases
no
isy
data
ca
n
be
sim
ply
id
entifi
ed
if
a
s
m
al
l
fr
act
ion
of
the
total
par
t
is
m
issi
ng
.
R
efer
to
“R
-
Squa
re”
wi
th
a
low
po
rtio
n
of
nois
y
data,
the
regressi
on
te
ch
nique
he
lps
div
i
de
ra
ndom
l
y
and
visua
li
zes
for
how
m
ajo
r
it
y
of
t
he
dete
rm
inistic
par
t
can
play
r
ole
in
the
fi
nal
outc
om
e.
No
te
t
hat
re
gressi
on
-
base
d
i
m
pu
ta
ti
on
is
a
lso
prese
nted
i
n
[
13,
14]
to f
igure o
ut
m
issin
g
value
s
an
d
t
he
ac
cu
racy
is m
easur
ed
by
ut
il
iz
ing
MSE
m
et
ric,
ho
we
ve
r,
it
is
r
egardless
of
es
tim
a
ti
on
m
od
el
.
No
isy
data
can
al
so
occ
ur
widely
as
bu
l
k
in
the
dataset
. I
n
this
case, the
m
ino
r
it
y part is f
resh
v
al
ue
an
d
is useab
le
, while
t
he
m
ajo
rity
o
ne
h
as to b
e d
isc
ard
e
d
or
retreat
e
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Esti
ma
ti
on
of r
egressi
on
-
base
d model w
it
h b
ulk noisy
data
(
Cha
ni
nto
r
n
Ji
tt
awiri
yan
uk
oo
n
)
3651
Howe
ver,
the
inv
e
sti
gation
i
n
this
pap
e
r
is
not
li
ke
oth
e
r
m
et
hodo
l
og
ie
s
a
s
m
entioned
a
bove
,
firstly
,
the
pr
opos
e
d
tr
eatm
ent
is
e
m
plo
ye
d
un
ti
l
the
bu
l
k
no
ise
is
unin
vo
l
ved.
Af
t
er
the
el
i
m
inatio
n
of
t
he
unwa
nted
bu
l
k
noisy
dat
a,
the
pro
pose
d
al
gorithm
keep
s
fi
xing
al
l
su
bst
antia
l
blank
s
i
n
the
data
set
by
est
i
m
a
tin
g
t
he
su
bst
it
ution
.
N
ote
that
the
si
ng
le
el
em
ent
of
noise
in
the
d
at
aset
can
ha
m
per
the
data
m
anipu
la
ti
on
un
le
ss
exclusi
on
of
t
his
m
ajo
r
nois
y
data.
F
our
pro
posed
treat
m
ents
are
int
r
oduce
d,
nam
ely,
D
up
li
cat
ion
(DU
),
Me
an
Va
riable
s
(MV
),
Delet
ion
Me
c
ha
nism
(D
M)
,
an
d
Ra
ndom
I
m
pu
ta
ti
on
(RI
).
T
hese
four
al
gorith
m
s
are
app
li
e
d
f
or
re
pl
aci
ng
no
ise
wi
th
substi
tuti
on.
Seco
ndly
,
the
com
pu
ta
ti
on
c
os
ts
ar
e
incl
us
ive
of
searc
hing
ti
m
e
for
the
el
im
inati
on
of
bulk
no
isy
data
and
in
div
id
ual
al
gori
thm
execu
ti
on
tim
e.
As
a
down
-
to
-
eart
h,
th
irdly,
these
fou
r
est
im
at
ion
s
are
val
idate
d
us
i
ng
t
he
com
par
ison
w
it
h
act
ual
val
ues
in
or
der
to r
eflect
their
ac
cur
acy
.
Last
ly
,
the
e
xp
e
rim
ental
resu
lt
s
us
i
ng
MOA
sim
ulati
on
a
re
s
umm
arized.
T
he
await
ing
sec
ti
on
s
a
re
cat
egorized
as
fo
ll
ows:
Sect
ion
III
ex
plains
the
existi
ng
bu
l
k
noise
s
it
uations
.
S
ect
ion
I
V
des
cribe
s
the p
e
rfo
rm
ance r
es
ults o
f
the
se pro
pose
d
te
chn
i
qu
e
s. Sec
ti
on V o
utli
n
es t
he
c
on
cl
us
io
n of t
he pa
per
.
3.
DA
T
AS
ET
S
WITH B
ULK
N
OI
SE
In
this
sect
ion,
featur
es
of
bu
lk
no
ise
are
il
lustrate
d,
datas
et
s
with
m
ajo
r
it
y
of
no
ise
ar
e
li
ste
d
and
the in
her
e
nt str
uctu
re
of
bulk
no
ise
is s
or
te
d
ou
t.
3.1
.
No
ise
pre
veni
ence
A
sin
gle
entry
of
no
ise
ca
n
corrupt
a
w
ho
le
dataset
.
No
i
se
can
be
ea
sil
y
or
iginate
d
f
ro
m
fau
lt
s
durin
g
data
c
ollec
ti
on
proce
dure
or
st
or
in
g
process
.
A
n
el
e
m
entary
existe
nce
of
noise
s
top
s
t
he
e
xtrac
ti
on
of
insig
htfu
l
data
in
data
c
ur
at
io
n,
wh
ic
h
ca
n
r
eflect
the
de
fe
ct
ed
m
achine
le
arn
i
ng
operati
on.
I
n
pr
act
ic
e,
it
can
be
des
per
at
el
y
com
plica
te
d
to
handle
fa
ults
as
s
uc
h.
The
r
efore,
t
o
pinp
oi
nt
an
d
t
o
treat
noise
in
a
ny
dataset
m
ay
con
quer
t
he
m
anipu
la
ti
on
co
ns
trai
nt
an
d
the
im
ped
i
m
ent
le
arn
i
ng
process.
In
t
his
pa
per,
th
us
,
the
severe
case
of
bulk
noise
in
the
dat
aset
is
inv
est
igate
d.
Bulk
no
i
se
turns
the
pr
esence
of
nois
e
in
the
dataset
to
be
beyo
nd
50%
of
the who
le
.
T
he
com
plexity
is
to
al
loca
te
an
d
searc
h
t
horou
gh
ly
w
he
re
t
he
bulk
noise
p
re
sents.
This
pinpo
i
nt c
an pr
ovide t
he c
on
cl
us
io
n reg
ard
i
ng the
nece
ssary of t
he bid
of
no
ise
t
reatm
ents.
To
pu
rg
e
bu
l
k
no
ise
,
m
os
t
te
chn
i
qu
e
s
assu
m
e
a)
the
dataset
is
ex
act
ly
at
hand
f
or
e
xe
cution
a
nd
b)
t
he
dataset
i
s
sm
a
ll
fo
r
m
a
nipulat
ing
at
a
tim
e.
This
is
due
to
t
he
im
ple
m
entat
ion
of
a
ru
le
-
base
d
a
ppro
ac
h.
But
in
this
pa
pe
r,
a
par
ti
ti
on
-
base
d
strat
egy
(P
BS)
is
ad
opte
d
by
giv
i
ng
that
a
dataset
S
can
be
div
ide
d
int
o
two
portio
ns
:
a
m
ino
r
pa
rt
(subset)
an
d
a
m
a
j
or
(bul
k)
pa
rt,
in
w
hich
t
he
bulky
par
t
is
ass
um
ed
to
prom
pt
th
e
identific
at
ion
of
the
m
ajo
r
no
i
se
from
the
whole
dataset
.
Co
ns
e
qu
e
ntly
,
the
assum
ption
is
reali
sti
c,
espec
ia
ll
y
in
no
isy
en
vi
r
on
m
ents.
In
ca
se
of
huge
dataset
s,
bu
l
k
noise
el
i
m
inati
on
is
scal
ing
up
pa
rtit
ion
filt
erin
g
tim
e
accor
dingly
.
T
he
ex
per
im
ent
on
synt
hetic
dataset
s
with
no
ise
le
vel
(e
ve
n
ab
ov
e
50%
)
shows
t
he
ef
fecti
ve
perform
ance aft
er P
BS.
A
m
od
est
te
ch
nique
[
15]
to
inv
olv
e
bulky
values
of
nois
y
data
is
to
re
m
ov
e
the
w
hole
instance
wh
ic
h
sp
eci
fic
al
ly
con
ta
ins
the
noise
data
of
any
at
trib
utes.
Howe
ver,
this
appro
ac
h
do
e
s
no
t
crac
k
the
pro
blem
in
case
of
bulk
no
ise
as,
on
ly
a
few
or
m
ino
rity
of
these
instance
s
rem
a
ins.
It
is
feasible
that
these
discar
ded
i
ns
ta
nces
m
ay
be
c
riti
cal
and
sen
sit
ive
factor
as
execu
ti
ng
data
cur
at
io
n.
T
o
identify
noise
in
the
dataset
is
desc
r
ibed.
Two
case
s
are
no
w
co
nsi
der
e
d,
the
ca
s
e
wh
e
re
the no
ise
is
rando
m
and
t
he
case
in w
hic
h
a
boun
d
on
th
e
noise
e
xists.
In
the
first
one,
i
n
case
of
huge
dataset
s,
a
synthesiz
e
r
will
com
pu
te
f
easi
ble
represe
ntati
on
s
fo
r
a
dataset
then
est
im
a
tes
the
dataset
base
d
upon
th
e
cal
culat
ion
.
In
the
la
tt
er
on
e
,
op
ti
m
iz
ation
is
expect
ed o
n
t
he
sim
ulatio
n.
3.2
.
Simul
a
ti
on
data
se
ts
Fil
te
ring
the
n
detachin
g
insta
nces
w
hich
are
m
ajo
rity
in
bulk
no
ise
datase
t
is
a
m
ai
n
pu
r
po
s
e
of
data
cur
at
io
n
pre
-
processin
g
(
bo
t
h
PBS
filt
ering
and
an
om
al
y
detect
ion
)
beca
us
e
t
hese
bu
l
k
noise
s
obstruc
t
then
stop
data
a
naly
ti
cs.
Pr
op
os
e
d
f
ou
r
t
reatm
e
nt
ap
proac
hes
for
est
im
a
ti
ng
data
f
or
m
ajo
rity
of
no
isy
values
wh
ic
h
are
D
upli
cat
ion
(
DU)
,
Me
an
Var
ia
bl
es
(MV)
,
Delet
ion
Me
cha
nism
(D
M),
an
d
Ra
ndom
I
m
p
utati
on
(RI)
ha
ve
bee
n
intr
oduce
d.
Let
T
be
a
give
n
noisy
datas
et
m
at
rix
wh
ic
h
co
ntains
a
r
ow
s
a
nd
b
c
olu
m
ns
,
wh
il
e
k
re
prese
nts
instan
ces
a
ff
ect
ed
by
bu
l
k
no
ise
or
noise
le
vel,
in
wh
ic
h
k
is
al
ways
l
ess
than
a
(
k
<
a
a
nd
T
k1
,
T
k2
,
T
k3
,…
,
T
k(
b
-
1)
,
T
kb
)
fo
r
eac
h
k
=
1,
2,
3,
…
,
a
.
T
he
T
m
at
rix
is
exp
e
ct
e
d
to
be
a
determ
inis
ti
c
set
.
An
el
em
ent
T
kb
is
set
to
be
a
no
isy
el
em
ent
wh
e
ne
ver
{
T
ij
=
ɸ
||
∞,
1
≤
i
≤
a
;
1
≤
j
≤
b
}
.
N
ote
that
in
c
ase
of
bu
l
ky
noise
,
k
is
al
ways
ident
ic
al
or
la
rg
e
r
t
han
t
he
hal
f
of
a
(
k
≥
a
/
2)
.
T
he
dataset
with
bu
l
k
noise
is
cal
led
ham
per
ed
dataset
.
Th
us
trea
t
m
ent
te
chn
iq
ues
t
o
ov
e
rth
r
ow
t
he
haza
rd
and
w
ork
ou
t
with
the
an
al
ysi
s
by apply
ing t
he
estim
at
ed
ve
ct
or
E
n
a
re
des
cribe
d
in
the
ne
xt secti
on.
The
par
ti
ti
on
-
base
d
filt
erin
g
te
chn
i
qu
es
e
m
ph
asi
ze
on
di
scard
in
g
no
is
e
wh
ic
h
ca
n
be
detect
ed
at
low
-
le
vel
data
fau
lt
s
est
ablis
he
d
by
a
n
im
pai
red
c
ollec
ti
on
of
data,
bu
t
i
nst
ances
w
hich
a
re
de
fected
ca
n
ba
r
the
analy
ti
cs.
No
ise
ca
n
di
re
ct
to
m
isi
nterp
ret
(n
e
gative
c
on
ce
r
n)
,
le
a
ding
data
a
naly
zer
to
ta
ke
on
a
r
el
at
ion
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19 :
3649
-
3656
3652
of
any
at
tri
bu
t
es
(
fau
lt
decisi
on)
al
th
ough
it
le
gitim
at
ely
is
not
(a
ty
pe
I
error).
The
refo
re,
if
it
is
to
c
ertai
n
data
analy
ti
cs,
these
T
kb
insta
nces
m
us
t
be
de
no
ise
d,
reg
a
r
dless
of
a
ny
an
al
ys
is.
It
is
al
so
crit
ic
al
for
de
no
is
e
appr
oach
es
to
detach
a
ny
noi
se
data.
In
cas
e
of
t
he
bu
l
k
portio
n
of
noise
wh
e
re
k
≥
a
/2,
any
m
et
ho
ds
hav
e
to
neg
le
ct
a
rem
a
ining
f
racti
on
of
t
he
w
ho
le
da
ta
set
as
a
rem
ov
al
ca
n
res
ult
a
certai
n
fe
w
or
only
m
ino
r
portio
n.
This
pa
pe
r
f
oc
us
es
on
f
our
ty
pes
of
the
pro
pose
d
treat
m
ent
for
bulk
noise
,
w
hich
incl
ude
Duplic
at
ion
(
DU),
Me
an
Va
riabl
es
(MV),
Del
et
ion
Me
cha
ni
sm
(D
M),
a
nd
Ra
ndom
I
m
pu
ta
ti
on
(R
I)
i
n
order
to
gran
t
data
analy
ti
cs
in
reali
ty
.
The
exp
e
rim
ents
are
based
on
th
e
re
gressi
ve
m
od
el
with
te
n
dif
fere
nt
syntheti
c
da
ta
set
s.
In
the
in
div
i
dual
exp
erim
ent,
the
est
i
m
ation
is
co
m
pu
te
d
a
fter
de
noisi
ng
in
te
rm
s
of
the
eff
ect
on
the
data
analy
sis.
I
n
the
co
ns
eq
ue
nt
ye
ar,
on
ce
the
ac
tual
data
is
c
ollec
te
d
then
,
th
ey
will
be
c
om
par
e
d
to
th
os
e
earl
y
ye
ar esti
m
at
ion
s.
4.
RESU
LT
S
A
ND AN
ALYSIS
The
open
-
s
our
ce
base
d
M
O
A
is
sel
ect
ed
for
the
data
a
naly
tic
s.
T
en
dataset
s
are
ta
ken
up
a
nd
the
assessm
ent
of
a
re
gr
e
ss
ion
m
od
el
for
bulk
noisy
with
the
dissi
m
il
ar
no
ise
le
vel
(
k
)
i
s
m
easur
e
d
.
The
ex
per
im
ent
is
m
anipu
la
te
d
on
an
I
ntel®
Core
™
i5
CP
U,
1.6
0
G
Hz
P
ro
ces
sor
an
d
8
GB
RAM
on
boar
d.
The dat
aset
s
a
r
e
ch
os
e
n
in
orde
r
that t
hey
al
l
are
di
ssim
il
ar
in
siz
e,
num
ber
of instance
s
,
a
nd att
rib
utes.
4.
1.
Mean
sq
uare
e
rror
Me
an
s
qu
a
red
error
(MSE
)
is
a
m
et
ric
to
qu
antify
t
he
di
f
fer
e
nces
betwe
en
sam
ple
and
popula
ti
on
values
a
ntici
pa
te
d
by
a
regressio
n
li
ne
or
f
or
eca
ste
d
va
lues
of
the
ob
s
er
vations.
The
lo
we
r
the
MSE,
the
cl
os
er
to
t
he
li
ne
of
bes
t
fit
is
fo
un
d.
The
MSE
ex
plain
s
the
sta
ndar
d
de
viati
on
of
the
dissim
il
arit
y
betwee
n
obser
vation
a
nd
pr
e
dicti
on
. T
he
di
ssi
m
il
ar v
al
ue
i
s calc
ulate
d
by the targ
et
ed
da
ta
ex
ecuti
on
over the
error
s
in
e
stim
at
ion
.
M
SE
bas
is i
s a b
al
a
nce
betwee
n varia
nc
e an
d bias is s
how
n
in
(
1
)
.
MSE
=
∑
(
̂
−
)
2
=
1
(
1
)
Wh
e
re,
is
tim
e
series
of
the
fi
nite
obse
rv
at
i
on,
̂
is
the
f
orec
ast
ed
ti
m
e
series,
a
nd
D
de
note
s
the
nu
m
ber
of
sam
ple
data.
A
dataset
in
a
trai
ni
ng
c
onditi
on
de
cl
ines
the
er
r
or
rate
f
or
e
xp
erim
ent
set
.
Fault
rate
for
trai
ni
ng
da
ta
set
will
be
com
par
at
ively
higher
t
han
th
at
of
t
he
ex
pe
rim
ent
set
.
If
any
tw
o
te
ch
ni
qu
es
pro
vid
e
t
he
e
qual
m
ean ab
s
olu
te
er
ror
t
hen
MSE i
s take
n u
p for
decidin
g,
wh
ic
h
is
t
he opt
i
m
u
m
app
r
oac
h
.
4.
2
.
Mean
ab
s
olute
error
The
m
ean
abs
olu
te
er
ror
(M
AE)
is
a
fig
ure
us
ed
t
o
qua
ntify
pr
e
dicti
on
s
of
the
c
riti
cal
resu
lt
s
.
The
M
AE
is a
m
ean
of the
absolute
value
of
fau
lt
s a
nd can
be defi
ned
by
MAE
=
1
∑
|
−
̂
|
=
1
(2)
Wh
e
re,
is t
he
f
init
e o
bse
r
vation t
i
m
e series
and
̂
is t
he pre
di
ct
ed
tim
e series.
4.3
.
No
ise
st
ru
ctur
es
Fu
ll
y
at
Ra
ndom
(F
AR)
def
i
ne
s
bulk
no
ise
s
tructu
res
a
re
not
relat
ing
t
o
a
ny
facto
rs.
For
exam
ple,
m
os
t
qu
est
io
ns
ask
the
res
po
nd
e
nt
for
a
ra
ndom
answ
e
r.
In
te
ntion
al
ly
(
ITT)
ou
tl
ines
bu
l
k
no
ise
str
uc
tures
add
e
d
up
with
pr
i
vacies.
F
or
instance,
s
om
e
respo
nd
e
nts
cl
um
si
ly
echo
their
sen
sit
ive
data
su
c
h
as
age
,
incom
e,
age,
et
c.
They
e
nd u
p
fill
ing
w
it
h
a b
la
nk d
el
iberat
el
y
or
w
hite
-
fa
lse
fig
ur
es
.
Ma
ny
bulk noise d
at
a
in
this pa
per incl
ud
e
both
I
T
T a
nd F
AR.
4.
4
.
Prop
os
e
d
tre
atmen
ts
Four
pro
po
se
d
treatm
ents
fo
r
assigni
ng
data
to
involve
t
he
prob
le
m
of
bu
lk
noisy
values
wh
ic
h
ar
e
base
d
on
duplica
ti
on
,
m
ean
value,
delet
ion,
and
ra
ndom
value
are
descr
i
be
d.
Let
T
be
a
giv
e
n
noisy
d
at
aset
m
at
rix
w
hich
c
on
ta
in
s
a
rows
an
d
b
c
ol
um
ns
,
wh
e
re
k
is
noise
le
vel
as
w
el
l
as
sm
aller
than
a
(
k
<
a
an
d
T
k
1
,
T
k
2
,
T
k
3
,…,
T
k
(
b
-
1)
,
T
k
b
)
fo
r
eac
h
k
=
1
,
2
,
3
,
…,
a.
The
T
matr
ix
is
e
xpect
ed
to
be
a
determ
inist
ic
set
.
An
el
em
ent
T
kb
is
set
to
be
a
no
isy
el
em
ent
wh
e
ne
ver
{
T
ij
=
ɸ
||
∞
,
1
≤
i
≤
a;
1
≤
j
≤
b
}
.
I
n
the
case
of
bu
lk
y
no
ise
,
k
is
e
qua
l
or
gr
eat
e
r
than
the
half
of
a
(
k
≥
a
/
2).
T
he
T
dat
a
set
is
cal
le
d
ham
per
ed
da
ta
set
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Esti
ma
ti
on
of r
egressi
on
-
base
d model w
it
h b
ulk noisy
data
(
Cha
ni
nto
r
n
Ji
tt
awiri
yan
uk
oo
n
)
3653
Th
us
,
treat
m
ent
in
or
de
r
to
ge
t
ov
er
the
st
oppag
e
a
nd
m
ov
e
on
with
the
a
naly
ti
cs
us
ing
the
est
i
m
at
ed
vecto
r
E
n
are
presente
d
as
f
ollows.
4.4.1
.
Dupl
ic
ati
on
(
DU)
Duplic
at
ion
in
vo
l
ve
s
with
t
he
dataset
T
by
firstly
delet
in
g
al
l
k
insta
nces
.
T
his
a
pproac
h
benefit
s
the
si
m
plest
treatm
ent
w
hethe
r
or
n
ot
the
re
are
i
m
pacts
on
the
el
i
m
inati
on
.
A
fter
the
entire
k
rows
of
the
m
at
rix
T
are
re
m
ov
e
d,
t
hen
est
im
a
te
d
E
n
data
set
is
{
d
ij
≠
ɸ
||
∞
,
1
≤
i
≤
(a
-
k)
;
1
≤
j
≤
b
}
.
Althou
gh
it
seem
s
the
D
U
treatm
ent
nurtur
es
a
n
unfair
est
i
m
ation
by
us
in
g
the
rem
ai
nin
g
a
-
k
inst
ances
to
dupli
cat
e
un
ti
l
En
datase
t
grow
s
to
{
d
ij
≠
ɸ
||
∞,
1
≤
i
≤
a;
1
≤
j
≤
b
}
but
the
paper
in
vestigat
es
a
bu
l
k
noise
sam
ple
and
ex
a
m
ines
wh
et
her o
r no
t
the DU a
ppr
oa
ch reflect
s
a
n
a
ccepta
ble st
rategy.
4.4.2
.
Me
an
vari
ab
le
s
(
MV)
Me
an
value
c
r
it
erion
is
to
im
pu
te
data
to
su
bst
it
ute
al
l
k
instances.
Apply
PBS
to
th
e
ta
rg
et
ed
T
datase
t
an
d
cl
a
ssify
a
dataset
wh
ic
h
c
om
pr
ise
s
of
k
instanc
es.
An
y
k
r
ows
of
the
m
at
rix
T
posse
ss
a
n
el
e
m
ent
d
ij
with
noisy
data
wh
e
re
{
d
ij
=
ɸ
||
∞
,
1
≤
i
≤
k;
1
≤
j
≤
b
}
then
the
entire
row
is
rep
la
ced
by
e
m
plo
yi
ng
the
MV s
ub
sti
tuti
on
for
est
im
at
ed
E
n
dataset
as li
ste
d
in
Equati
on
3.
=
1
|
−
|
∑
=
+
1
(3
)
The
in
vestiga
ti
on
of
the
MV
is
that
it
is
an
acce
ptable
est
im
a
ti
on
for
a
pa
ram
et
er
ou
t
of
a
norm
al
distrib
ution.
In
case
o
f
I
TT,
this
treat
m
ent
induces
a
vo
la
ti
le
bias.
No
t
to
m
ention
t
he
MV
is
le
d
by
the
dist
or
te
d repla
ce
m
ent
as w
el
l
as
dev
el
op
s
th
e size
of stat
e s
pace
c
om
par
ed
to
the
abo
ve D
U
.
4.4.3
.
Deleti
on
me
ch
an
ism
(
DM)
DM
is
si
m
ply
rem
ov
ing
al
l
k
instances
.
T
he
D
M
cl
aim
s
the
si
m
ple
bu
t
qu
i
cker
treat
m
ent
w
het
her
or
no
t
t
he
delet
io
n
will
influ
e
nc
e
the
fu
t
ur
e
a
na
ly
ti
cs
.
Af
te
r
t
he
e
ntire
k
rows
of
the
m
at
rix
T
a
re
delet
ed,
t
he
n
est
i
m
at
ed
E
n
datase
t
is
{
d
ij
≠
ɸ
||
∞
,
1
≤
i
≤
(a
-
k)
;
1
≤
j
≤
b
}
.
The
DM
treat
m
ent
pr
om
ote
s
hu
m
ble
analy
ti
cs
and
a
fair
pr
e
dicti
on
in
wh
ic
h
s
ta
te
sp
ace
is
insig
nificant
,
there
fore,
t
he
DM
is
a
provoki
ng
appr
oach.
4.4.4
.
R
an
d
om
im
puta
tion
(R
I
)
Util
iz
e
sever
al
i
m
pu
ta
ti
on
s
a
t
rando
m
fo
r
r
eplace
m
ent.
Sim
il
ar
to
MV,
the
PBS
is
ap
plied
to
t
he
ta
rg
et
ed
T
data
set
an
d
resu
lt
s
a
da
ta
set
with
k
insta
nces.
A
ny
k
rows
of
t
he
m
at
rix
T
possess
a
n
el
em
e
nt
d
ij
with
noisy
dat
a
wh
e
re
{
d
ij
=
ɸ
||
∞,
1
≤
i
≤
k;
1
≤
j
≤
b
}
t
hen
t
he
e
ntire
row
is
s
ub
sti
tu
te
d
by
us
in
g
t
he
RI
rep
la
cem
ent f
or esti
m
a
te
d
E
n
dataset
.
The
m
ini
m
u
m
li
kelihood
f
ound
in
c
olu
m
n
j
(w
he
re
j
=
1
,
2
,
3
,…
,b
)
is
m
ark
ed
by
d
(
m
in)
j
wh
e
r
e
d
(m
in)
j
=
Mi
n
(
d
kj
)
for
each
k
=
1
,
2
,
3
,…
,
(
a
-
k)
.
Likewise,
t
he
m
axi
m
u
m
l
i
kelihoo
d
of
col
um
n
j
(w
her
e
j
=
1
,
2
,
3
,…,
b
)
is
de
fine
d
by
d
(m
ax)
j
w
he
re
d
(m
ax)
j
=
Ma
x
(
d
kj
)
fo
r
eac
h
k
=
1
,
2
,
3
,…,
(
a
-
k)
.
The
substi
tuti
on
for
est
i
m
at
ed
E
n
d
at
aset
with
m
ulti
ple
i
m
pu
ta
tio
ns
for
k
instances
in
each
c
olu
m
n
j
is
rando
m
ly
exp
la
ined
as
fo
ll
ows:
=
[
(
)
,
(
)
]
(4)
Stat
e
sp
ace
can
be
sig
nifi
cant
to
ref
le
ct
the
sp
eed
of
com
pu
ti
ng
pow
e
r.
In
this
pap
e
r
,
the
com
pu
ta
ti
on
c
os
t
is
sum
m
arized,
accor
ding
to
the
perform
ance
evaluati
on.
It
is
app
a
ren
t
th
at
any
pr
e
dicti
on
s
are
pro
blem
atical
if
the
c
om
pu
ta
ti
on
c
os
t
is
hi
gh
as
de
picte
d
i
n
Ta
ble
1.
N
ote
that
in
case
of
bu
l
k
no
ise
,
a
is al
w
ay
s sm
a
ll
er th
an 2
k
.
Table
1.
C
om
pu
ta
ti
on
c
os
t
of
pro
po
se
d
t
reatm
ent
s
Tr
eat
m
en
t
Co
m
p
u
tatio
n
Co
st
DU
O(
ab
)
+
O(
ab
-
bk
)
≈ O(
ab
)
MV
O(
ab
)
+
O(
ab
-
bk
)
≈ O(
ab
)
DM
O(
ab
)
RI
O(
ab
)
+
O(2(
ab
-
bk
))
≈
O(
ab
)
It
is
propose
d
that
treatm
ents
fo
r
i
nvolv
i
ng
the
bulk
no
ise
can
be
ou
tl
ined
into
tw
o
colle
ct
ive
strat
egies,
a
m
od
el
-
base
d
stra
te
gy
(MBS),
a
nd
a
par
ti
ti
on
-
base
d
strat
egy
(P
BS)
.
Th
e
PBS
will
div
ide
a
nd
ir
on
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19 :
3649
-
3656
3654
ou
t
t
he
bu
l
ky
no
ise
pa
rt
be
f
or
e
ap
plyi
ng
the
est
im
a
t
ion
.
Wh
il
e
th
e
MB
S
re
vises
t
he
a
lgorit
hm
s
in
orde
r
t
o
carry
ou
t
the
noise
data
befo
r
e
the
pa
ram
et
ric
est
i
m
ation
is
ap
plied.
T
he
com
m
on
MB
S
is
us
e
d
in
P
SPP
or
ANO
VA
s
of
t
w
are,
w
hich
a
ppli
es
var
ious
i
m
pu
ta
ti
ons
f
or
r
eplaci
ng
a
no
is
e
data.
PBS
te
chn
i
qu
e
im
ple
m
ents
li
kelihood
in
form
ation
into
at
te
ntion
rat
her.
W
it
h
the
MB
S
al
go
rithm
,
the
e
m
plo
ym
ent
i
s
com
plex
and
the
sk
il
l i
s r
eq
uire
d because
the a
lgorit
hm
p
er s
e
h
as
bee
n
f
undam
ental
ly
d
esi
gn
e
d
t
o
re
flect
the p
a
ram
et
er cen
tric
especial
ly
to
the
sta
te
sp
aces
.
If
t
he
MB
S’
s
al
go
rit
hm
pr
ovides
c
omm
on
resu
lt
s,
s
uc
h
as
a
var
ia
nce
or
vector
of
a
ve
rag
e
val
ues,
t
hen
it
can
be
cal
le
d
th
e
data
ce
ntric.
The
e
rror
m
etr
ic
s
of
te
n
dis
si
m
il
ar
dataset
s
us
i
ng
MOA
at
no
ise
le
vel
ran
ging
fr
om
50
%
to
80
%
will
be
colle
ct
ed.
Th
is
is
a
si
m
ple
analy
sis
towar
d
the
desig
nated
dat
aset
s,
an
d
res
ul
ts
are
li
ste
d
in
Ta
ble
2
-
5
.
T
he
th
ree
er
rors
in
the
ta
ble
c
har
act
erize
t
he
m
ea
n
sq
ua
re
d
er
r
or
(
MSE),
t
he
c
orrelat
ion
c
oeffi
ci
ent
(COE
F),
and
m
ean
absol
ute
error
(M
A
E)
c
orrespo
nd
i
ng
ly
.
Dataset
#5
pro
v
ides
l
owest
a
m
ou
nt
f
or
both
MA
E
a
nd
MSE
w
hile
da
ta
set
#2
,
#7,
#8,
a
nd
#9
obta
in
s
m
al
le
st
fig
ur
e
f
or
C
OE
F.
T
he reg
ressi
on
-
based p
re
dicti
on
is als
o
s
um
m
arized in
T
able 6.
Table
2.
E
stim
at
ion
with
m
ean s
quare
d
e
rror
for
te
n di
ff
e
rent
d
at
aset
s
(
k
=
0.5)
Dataset
MSE
COEF
MAE
1
1
5
.7
0
.47
1
3
.1
2
1
.8
0
.1
1
.6
3
30
0
.3
2
5
.7
4
2
7
.8
0
.15
2
3
.5
5
0
.12
0
.4
0
.1
6
2
.55
0
.6
1
.8
7
2
2
.3
0
.64
1
9
.16
8
1
3
6
.1
0
1
1
4
.8
9
5
0
.02
3
.77
10
1
7
.68
0
.02
1
3
.4
Table
3.
E
stim
at
ion
with
m
ean s
quare
d
e
rror
for
te
n di
ff
e
rent
dataset
s
(
k
=
0.6)
Dataset
MSE
COEF
MAE
1
1
6
.7
0
.63
1
3
.9
2
1
.59
0
.27
1
.34
3
2
9
.2
0
.47
2
5
.6
4
2
8
.2
0
.16
24
5
0
.12
0
.42
0
.1
6
2
.6
0
.5
1
.9
7
2
5
.3
0
.64
21
8
1
3
7
.55
0
.15
1
1
6
.1
9
5
0
4
.1
10
1
2
.8
0
.5
9
.5
Table
4.
E
stim
at
ion
with
m
ean s
quare
d
e
rror
for
te
n di
ff
e
rent
d
at
aset
s
(
k
=
0.7)
Dataset
MSE
COEF
MAE
1
1
9
.2
0
.53
1
5
.6
2
1
.78
0
.2
1
.48
3
3
3
.3
0
.38
2
9
.7
4
2
8
.8
0
.28
25
5
0
.12
0
.4
0
.1
6
2
.86
0
.44
2
7
3
2
.2
0
.48
27
8
1
2
7
.8
0
.29
102
9
5
.8
0
.1
4
.8
10
1
8
.41
0
.14
1
3
.14
Table
5.
E
stim
at
ion
with
m
ean s
quare
d
e
rror
for
te
n di
ff
e
rent
d
at
aset
s
(
k
=
0.8)
Dataset
MSE
COEF
MAE
1
2
4
.2
0
.25
2
0
.5
2
1
.75
0
.11
1
.55
3
4
4
.1
0
.22
3
8
.5
4
2
5
.3
0
.57
2
1
.1
5
0
.12
0
.36
0
.1
6
1
.7
0
.76
1
.24
7
4
6
.4
0
.11
3
2
.1
8
1
3
8
.5
0
.32
1
1
2
.4
9
5
.05
0
.8
4
10
4
.1
0
.92
1
.7
Table
6.
Re
gr
e
ssion
-
ba
sed
pr
edict
ion
f
or
te
n datase
ts
Reg
ressio
n
-
b
ased
Pr
ed
ictio
n
DATAS
ET
1
X
9
=
-
6
.07
X
4
+3
9
.85
2
X
11
= 0
.58
X
9
+1
.75X
5
+9
.4X
3
+2
.14
3
X
8
=
-
0
.91
X
3
+8
.2X
5
+ 2
1
6
.44
4
X
4
=0
.34
X
1
+2
.26
X
5
-
1
5
.18
5
X
2
=X
6
+0
.2X
1
+0
.48
6
X
6
= 0
.3X
3
+0
.1X
7
-
6
.96
7
X
5
=8
.63
X
1
+0
.8X
3
+ 1
1
.97
8
X
7
= 0
.09
X
5
+1
9
9
.1
1
9
X
3
=3
.8X
1
+3
4
.82
X
5
+X
4
+1
9
.7
10
X
6
=
-
4
.7X
1
-
4
.8X
2
+X
4
-
1
.4X
5
+0
.1X
7
+7
4
.2
Table
s
7
-
10
re
veal
aver
a
ge
e
rror
s
f
or
a
regressio
n
-
base
d
est
i
m
ation
co
m
par
ing
to
t
he
act
ual
data.
In
t
his
pa
pe
r,
t
en
diff
e
re
nt
da
ta
set
s
are
m
ea
su
re
d
a
nd
the
no
ise
le
vel
(
k
)
is
rangin
g
fro
m
50
%,
60%,
70%,
and
80%
as
ta
bu
la
rized
in
T
able
s
7
-
10
res
pecti
vely
.
The
resu
lt
s
in
m
o
st
cases
disp
la
y
the
est
i
m
at
ion
with
rand
om
i
m
pu
t
at
ion
(RI)
ca
n
m
ini
m
iz
e
the
a
ver
a
ge
e
rro
r.
I
n
a
dd
it
io
n,
in
case
of
bulk
noise
,
t
he
c
om
pu
ta
ti
on
cost
for
al
l
four
treat
m
ents
i
s
approxim
at
ely
identic
al
.
It
confirm
s
RI
is
the
m
os
t
eff
ect
ive
m
echan
ism
fo
r
bu
l
k no
ise
an
al
ysi
s.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
Esti
ma
ti
on
of r
egressi
on
-
base
d model w
it
h b
ulk noisy
data
(
Cha
ni
nto
r
n
Ji
tt
awiri
yan
uk
oo
n
)
3655
Table
7
.
A
ver
a
ge perce
ntage
of er
ror fo
r
te
n
di
ff
e
ren
t
da
ta
set
s
(
k
=0
.5)
k
= 0.5
Dataset
DU
MV
DM
RI
1
1
2
5
.8
1
1
3
.6
1
4
0
.2
1
0
5
.7
2
7
3
.65
7
4
.64
8
0
.17
7
2
.7
3
1
7
.3
1
7
.6
3
2
.8
1
6
.9
4
2
5
.8
2
5
.2
4
7
.3
2
6
.5
5
8
2
.3
5
0
.4
3
3
.4
2
3
.2
6
2
5
.7
2
2
.7
5
6
.1
2
5
.5
7
2
9
.03
2
7
.88
2
8
.57
2
5
.2
8
2
4
.25
2
4
.28
3
3
.5
2
5
.1
9
5
6
.6
5
6
.8
69
55
10
2
5
.6
2
2
.1
2
4
.9
4
4
.9
Table
8
.
A
ver
a
ge perce
ntage
of er
ror fo
r
te
n
di
ff
e
ren
t
da
ta
set
s
(
k
=0
.6)
k
= 0.6
Dataset
DU
MV
DM
RI
1
1
2
8
.8
114
1
4
6
.9
1
1
5
.7
2
7
4
.4
7
4
.2
8
0
.2
7
3
.3
3
1
7
.2
1
7
.7
3
6
.1
1
8
.3
4
2
5
.6
2
4
.9
46
2
4
.6
5
8
3
.2
4
4
.8
3
3
.5
2
3
.2
6
2
4
.7
23
63
2
5
.9
7
2
8
.7
2
7
.1
2
7
.8
2
4
.7
8
2
4
.4
2
4
.3
3
2
.1
2
4
.5
9
5
6
.3
5
6
.7
6
7
.9
5
4
.7
10
2
4
.8
2
1
.2
2
4
.2
5
0
.3
Table
9
.
A
ver
a
ge perce
ntage
of
er
ror fo
r
te
n
di
ff
e
ren
t
da
ta
set
s
(
k
=0
.7)
k
= 0.7
Dataset
DU
MV
DM
RI
1
1
3
3
.3
1
1
8
.5
1
5
5
.3
9
8
.8
2
7
3
.1
7
3
.3
7
9
.3
7
2
.3
3
1
6
.8
1
7
.7
41
1
7
.9
4
2
5
.2
2
4
.8
4
5
.7
2
2
.2
5
8
3
.7
4
0
.3
3
3
.4
2
3
.3
6
2
6
.4
2
4
.4
6
9
.4
2
5
.6
7
3
0
.6
2
8
.6
2
9
.3
2
5
.9
8
2
4
.6
2
4
.3
3
0
.7
2
5
.2
9
5
6
.1
5
6
.8
6
5
.4
5
4
.9
10
2
4
.9
2
1
.1
2
4
.2
4
4
.3
Table
10
.
Av
e
r
age
per
ce
ntage
of err
or
f
or
te
n
di
ff
e
ren
t
da
ta
set
s
(
k
=0
.8)
k
= 0.8
Dataset
DU
MV
DM
RI
1
1
5
1
.2
1
2
6
.8
1
8
9
.1
1
1
0
.6
2
6
9
.9
6
9
.6
7
9
.8
6
9
.4
3
1
6
.3
1
7
.3
4
5
.3
1
6
.7
4
2
5
.1
2
4
.9
4
7
.5
2
4
.3
5
8
4
.1
3
4
.6
3
5
.2
2
3
.4
6
27
2
4
.8
7
9
.6
2
7
.8
7
3
2
.7
2
9
.3
34
2
8
.1
8
2
4
.6
2
4
.1
2
8
.8
2
5
.1
9
5
5
.6
5
6
.6
6
2
.8
5
2
.9
10
2
5
.2
2
1
.6
2
4
.8
3
6
.3
5.
CONCL
US
I
O
N
In
t
his
pa
pe
r,
a
re
gr
essi
ve
-
base
d
est
im
a
tio
n
is
prese
nted
as
on
e
of
t
he
ef
fecti
ve
t
oo
ls
of
data
analy
ti
cs.
I
n
pract
ic
e,
the
an
al
yt
ic
al
data
after
c
ollec
ti
on
has
st
um
bled
upon
th
e
no
is
y
data
wh
ic
h
r
eflect
s
directl
y
to
con
t
ents
f
ound
in
instances
or
at
tr
ibu
te
s
in
the
da
ta
set
.
The
pro
blem
of
no
ise
values
is
c
omm
on
in
countle
ss
stu
di
es
and
can
disco
ntin
ue
th
e
rest
in
dat
a
cur
at
io
n.
It
has
bee
n
ex
te
ns
ively
ta
ke
n
int
o
consi
der
at
io
n
i
n
the
case
of
t
reatm
ents
so
t
hat
the
on
go
i
ng
analy
sis
can
be
fig
ured
ou
t.
Hen
ce
,
this
arti
cl
e
deep
ly
inv
e
sti
gates
the
influe
nce
of
bu
l
k
noise
data
and
t
he
pr
eci
si
on
of
ind
ivi
dual
treatm
ent
af
te
r
getti
ng
ov
e
r.
Te
n
diss
i
m
i
la
r
dataset
s
with
differe
nt
no
ise
le
vel
ar
e
ta
ken
up
to
est
i
m
at
e
by
us
ing
re
gr
e
ssive
base
d
m
od
el
s
against
the
act
ual
val
ues.
The
ave
ra
ge
per
ce
ntage
of
er
rors
from
fou
r
pro
po
se
d
treat
m
ents,
na
m
ely
DU,
MV,
DM,
an
d
RI
are
s
tud
ie
d
for
a
bulk
no
ise
filt
erin
g
ap
proac
h.
It
is
pr
ove
n
that
RI
dese
rv
e
s
th
e
identic
al
com
pu
ta
ti
on
c
os
t,
c
an
be
util
iz
ed
for
bu
l
k
no
ise
treat
m
ent,
and
ou
t
perform
s
than
ot
her
s
in
m
any
cases
for
the
est
i
m
ation
.
Dat
a
char
act
erist
ic
s
su
ch
as
co
va
riance
an
d
kurt
os
is
analy
sis
of
the
data
will
be
the
tria
l of
t
he fut
ure re
searc
h.
REFERE
NCE
S
[1]
M.
J.
B
asha
a
nd
K
.
P.
Kaliyam
urthi
e
,
"
An
Im
prove
d
Sim
il
ari
t
y
M
at
ch
ing
base
d
Clust
eri
n
g
Fram
ework
for
Short
and
Sent
e
nce
Le
ve
l
T
ext,
"
Inte
rnational
Journal
of
El
e
c
tric
al
and
Computer
Engi
n
ee
ri
ng
,
v
ol
.
7,
no.
1,
pp.
551
-
558
,
20
1
7
.
[2]
S.
J.
L
ee
and
J.
Y.
Jian
g,
"
Multi
la
b
el
Te
x
t
Categorizat
ion
Based
on
Fuz
z
y
Re
le
van
ce
Cluste
ring
,
"
IE
EE
Tr
ansacti
ons on Fuzz
y
Syste
ms
,
v
ol.
22
,
pp
.
1457
-
1471,
2014
.
[3]
V.
T.
W
ei
,
Y.
L
u,
H.
Ch
ang,
Q.
Zhou
and
X.
B
ao,
"A
sem
ant
i
c
appr
oa
ch
for
text
cl
ust
eri
ng
usi
ng
W
ordNet
an
d
le
xi
ca
l
ch
ai
ns
,
"
Ex
pert
Syste
ms
wit
h
App
licati
on
s
,
vol. 42, pp. 22
64
-
2275,
2015
.
[4]
C.
End
ers,
Appl
i
ed
Missing
Data
Analysis
,
New
York
:
Guilford
Press
,
2010.
[5]
G.
Li
u,
Y
.
W
ang
,
T.
Zha
o
,
and
D
.
Li,
"Resea
r
ch
o
n
the
par
al
l
el
t
ex
t
cl
uste
ring
a
lgor
it
hm
base
d
on
th
e
sem
ant
ic
tree
,
"
Proce
ed
ing
of
t
he
6th
Inte
rnat
i
onal
Confe
renc
e
on
Computer
Sci
en
ce
s
and
Co
nve
rgenc
e
Infor
mation
Technology
(
ICCIT
)
,
pp.
400
-
403,
201
1
.
[6]
K.
Shi
and
L.
Li
,
"H
igh
per
fo
rm
anc
e
gene
t
ic
al
gorit
hm
-
base
d
te
xt
cl
uster
ing
using
par
ts
of
s
pee
ch
and
out
lier
el
imination
,
"
Ap
pli
ed
Intelli
g
ence,
vo
l. 38, pp. 51
1
-
519,
2013
.
[7]
C.
L
i,
Y.
Ta
n
,
a
nd
J.
Kong,
"A
n
Maha
l
anobi
s
di
stanc
es
b
ase
d
text
c
luste
ring
a
lg
orit
hm
,
"
Int
ernati
onal
Con
fe
ren
c
e
on
Aut
omat
ic
Co
ntrol
and
Arti
f
ic
i
al
Int
el
l
ige
nc
e
(
ACA
I)
,
pp.
465
-
468
,
2012
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
9
, N
o.
5
,
Oct
ober
20
19 :
3649
-
3656
3656
[8]
S.
Bano,
K.
R.
Rao
and
E.
S
.
P
rak
ash,
"P
art
i
al
Conte
xt
Sim
il
ar
i
t
y
of
Gen
e/
Prot
ei
ns
in
Le
uk
emia
Us
ing
Conte
xt
Rank
Based
Hie
rar
chica
l
C
luste
r
ing
Algorit
hm
,
"
Inte
rnational
J
ournal
of
El
ectric
al
and
Computer
Engi
ne
erin
g
(
IJE
CE)
,
vol. 5
,
no.
3
,
pp
.
483
-
4
90,
201
5.
[9]
M.
A.
Shabir
and
P.
T.
Dee
pal
i
,
"
Sate
llite
Im
age
Denoising
Us
ing
Discre
te
Cosine
Tra
nsform
,"
Indone
sian
Journal
of
E
le
c
tric
al
En
gine
ering
and
In
formatic
s
(
IJE
EI
)
,
v
ol.
5
,
pp
.
3
72
-
375
,
2017
.
[10]
A.
Bife
t
,
R
.
Kirk
b
y
,
G.
Holm
es
a
nd
B.
Pfahring
er
,
"
MO
A:
Mass
ive
Online Analy
si
s
,
"
Journal
of
M
achi
ne
Learning
Re
search
,
vol
.
1
1,
pp
.
1601
-
160
4,
2010
.
[11]
G.
Madhu
,
e
t.
a
l.
,
"
A
Non
-
Para
m
et
ri
c
Discre
ti
z
at
i
on
Based
Im
puta
ti
on
Algor
it
hm
for
Conti
nuous
Attri
but
es
with
Miss
ing
Data
V
a
lue
s,
"
In
te
rnatio
nal
Journal
of
In
formation
Proc
e
ss
ing
,
vol. 8, no. 1, 64
-
72,
2014.
[12]
A
.
Purw
ar
and
S
.
K
.
Singh,
"
Hy
brid
p
red
i
ct
ion
m
odel
with
m
issing
val
ue
imp
uta
ti
on
for
m
edica
l
d
at
a
,
"
E
xpe
r
t
Syst
ems wi
th
Ap
p
li
cations
,
vo
l.
4
2,
no
.
13
,
pp.
562
1
-
5631
,
2015
.
[13]
G.
Madhu
and
G
.
Naga
ch
andr
ika,
"
A ne
w pa
rad
igm for
deve
lopm
ent
of
dat
a
imput
at
ion
appr
o
ac
h
f
or
m
issing va
lue
esti
m
at
ion
,
"
Inter
nati
onal
Journ
al
of
Elec
tric
al
and
Computer
Engi
nee
ring
(
IJECE)
,
vol.
6,
no
.
6,
pp.
3222
-
322
8,
2016.
[14]
R.
Thi
rum
aha
l
a
nd
Pati
l
A.
Dee
pal
i
,
"
KN
N
and
ARL
Based
Im
p
uta
ti
on
to
Est
imate
Miss
ing
Value
s
,
"
Indone
sian
Journal
of
Elec
t
rical
Engi
ne
erin
g
and
Informati
c
s
(
IJE
EI)
,
vol. 2
,
no.
3
,
pp.
119
-
1
24,
2014
.
[15]
Nee
raj
R
aheja
a
nd
V.
K.
Katiy
a
r,
"
Noise
red
uc
t
ion
appr
oa
ch
ba
sed
on
n
x
1
table
and
XS
L
di
spla
y
m
e
thod
fo
r
eff
icient web
d
ata
ex
tracti
on
,
"
Int
ernati
onal
Journ
al
of
Compute
r
Appl
ic
a
ti
ons
,
vo
l.
64
,
2013
.
Evaluation Warning : The document was created with Spire.PDF for Python.