Int
ern
at
i
onal
Journ
al of Inf
orm
at
ic
s
and
Co
m
munic
at
i
on
Tec
hn
olog
y (IJ
-
I
CT)
Vo
l.
6
,
No.
3
,
D
ece
m
ber
201
7
, pp.
199
~
208
IS
S
N:
22
52
-
8776
,
DOI: 10
.11
591/iji
ct
.
v6
i
3.p
p1
99
-
208
199
Journ
al h
om
e
page
:
http:
//
ia
esj
ou
r
nal.co
m/
on
li
ne/in
dex
.php
/
IJ
ICT
Density
Based Cl
usterin
g with Int
egrated
On
e
-
Class SVM
for
Noise Re
du
ctio
n
K. Nafees
Ahm
ed
*
,
D
r. T.
Ab
d
ul R
az
ak
*
Depa
rte
m
ent of
Com
pute
r
Sci
en
ce
,
Jam
al
Moh
a
m
ed
Coll
eg
e, Ta
m
il
nadu, I
ndi
a
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
A
ug
12
nd
,
20
1
7
Re
vised
Oct
26
th
, 201
7
Accepte
d
Nov
6
th
, 201
7
Inform
at
ion
ex
t
ra
ction
from
da
ta
is
on
e
of
th
e
ke
y
necess
it
i
es
for
da
ta
ana
l
y
sis.
Uns
up
erv
ised
na
ture
of
dat
a
l
ea
ds
t
o
complex
co
m
puta
ti
onal
m
et
hods
for
ana
l
y
sis.
Th
is
pape
r
pre
sents
a
density
b
ase
d
spatial
cl
usteri
ng
te
chn
ique
in
te
gr
at
ed
wi
th
one
-
cl
ass
Support
Vec
tor
Ma
chi
ne
(SV
M),
a
m
ac
hine
l
ea
rn
in
g
te
chn
ique
for
noise
re
du
ct
io
n,
a
m
odified
var
ia
n
t
of
DBS
CAN
ca
ll
e
d
Noise
Reduced
DBS
CAN
(
NRD
BS
C
AN
).
Anal
y
sis
of
DBS
CAN
exhi
b
it
s
it
s
m
aj
or
re
q
uire
m
ent
of
ac
c
ura
te
thre
sholds,
abse
nce
o
f
which
y
ie
lds
su
bopti
m
al
re
sults
.
How
eve
r,
id
e
nti
f
y
ing
a
cc
ur
ate
thre
shold
sett
ings i
s un
at
t
a
ina
bl
e.
Noise
is
one
of
th
e
m
aj
or
side
-
eff
ec
ts
of
t
he
thr
eshold
gap.
The
propo
sed
work
re
duc
es
noise
b
y
int
e
gra
ti
ng
a
m
ac
h
i
ne
learni
n
g
cl
assifi
er
in
to
th
e
oper
at
ion
struc
ture
of
DBS
CAN
.
The
Exp
eri
m
ent
a
l
re
sul
ts
indi
c
at
e
high
ho
m
ogene
ity
le
v
el
s in the c
lust
eri
ng
proc
ess.
Ke
yw
or
d:
Cl
us
te
rin
g
DBSCA
N
Ma
chine
Le
ar
ni
ng
Cl
assifi
e
r
No
ise
Red
uction
One
-
cl
ass
SVM
Copyright
©
201
7
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
:
K.
Nafee
s
Ahm
ed
,
Dep
a
rtem
ent o
f
Com
pu
te
r
Sci
ence
,
Jam
a
l M
oh
am
e
d
C
ollege
,
7
Ra
ce C
ourse
Road
,
Kh
a
j
a
N
agar
,
Tir
uc
hira
pp
al
li
, Tam
il
N
adu,
620
0
2
0
,
I
nd
ia
.
Em
a
il
:
naf
eesj
m
c
@
gm
ail.co
m
1.
INTROD
U
CTION
The
c
urren
t
I
nt
ern
et
a
ge
is
data
ric
h,
but
info
rma
ti
on
poor.
Me
anin
gful
da
ta
is
the
chief
r
equ
i
rem
ent
of
the
c
urre
nt
age.
T
his
has
le
d
to
an
e
nor
m
ou
s
increase
in
the
data
pr
ocessin
g
te
ch
ni
qu
es
.
H
ow
e
ve
r,
th
e
m
ajo
r
dr
a
wb
ac
k
is
inh
ere
ntly
e
m
bed
de
d
in
the
data
it
sel
f.
The
distri
bu
ti
on
of
d
at
aset
s
va
ry
and
he
nce a
sing
le
te
chn
iq
ue
that
was
desi
gn
e
d
to
process d
at
a
from
a
do
m
ai
n
would
not
po
s
sibly
pr
ovide
e
ff
ect
ive
res
ults
wh
e
n
app
li
ed
to
th
e
data
from
the
sa
m
e
do
m
ai
n,
du
e
to
va
riat
ions
in
the
data
distribu
ti
on
le
vel
s
as
tim
e
pr
ogr
esses.
Hen
ce
tech
niques
t
hat are a
s
du
ct
il
e as t
he d
at
a it
sel
f
are th
e only
ones t
ha
t can
per
sist
.
Un
s
uper
vise
d
data
proces
sin
g
has
al
ways
be
en
a
chall
en
ge
du
e
t
o
their
unpre
dicta
ble
na
ture.
Duct
il
e
and
tract
a
ble
al
go
rithm
s
are
the
m
ajo
r
requirem
ents
fo
r
processi
ng
s
uc
h
data.
D
om
ai
ns
with
su
c
h
requirem
ents
ran
ge
f
r
om
i
m
age
processi
ng
to
web
in
f
or
m
at
ion
proc
essing.
The
r
equ
i
rem
ents
fo
r
s
uch
processi
ng techn
i
qu
e
s ar
e
constantl
y o
n
t
he rai
se,
with th
e
increase i
n
t
he am
ou
nt
of
data b
ei
ng a
vaila
bl
e.
Extracti
ng
m
e
anin
gful
in
for
m
at
io
n
from
su
ch
data
r
eq
ui
res
gr
ouping
t
hem
to
find
c
omm
on
al
it
ies
existi
ng
bet
we
en
them
.
This
ai
ds
in
bette
r
i
nter
pr
et
at
io
n
of
data.
Cl
us
te
ri
ng
is
the
proce
ss
of
groupin
g
data
su
c
h
that
data
within
a
gro
up/c
luster
is
m
or
e
coh
esi
ve
c
om
par
ed
to
data
in
diff
ere
nt
cl
us
te
rs
.
The
pro
cess
of
cl
us
te
rin
g
m
akes
the
data
m
e
anin
gful
f
or
va
rio
us
ap
plica
ti
on
s
.
Cl
us
te
ri
ng
m
et
ho
ds
are
us
ually
cl
assif
ie
d
as
par
ti
ti
on
al
a
nd
hierarc
hical
c
lusterin
g
te
ch
ni
qu
es
.
Partit
io
nal
cl
us
te
ri
ng
te
chn
iq
ues
pe
r
form
flat
cl
us
t
erin
g
base
d
on
a
s
ing
le
decisi
on
cr
it
erion
su
c
h
as
distance
or
de
nsi
ty
.
Dista
nce
base
d
cl
us
te
rin
g
te
chn
i
ques
in
cl
ud
e
K
-
Me
a
ns
[
1]
cl
us
te
rin
g
an
d
CLARA
[
2]
to
nam
e
a
few
.
Den
sit
y
base
d
cl
us
te
rin
g
te
ch
niques
are
c
urr
ently
on
the
raise
du
e
t
o
their
fle
xib
le
op
e
rati
onal
nat
ur
e
a
nd
the
sta
ble
so
l
ution
set
s
ge
ner
at
e
d
by
them
.
Den
sit
y
base
d
cl
us
te
rin
g
te
ch
niques
incl
ude
DBSCA
N
[3]
and
De
nClu
e
[4
]
,
OPTIC
S
[
5]
to
nam
e
a
fe
w.
Hiera
r
chical
cl
us
te
rin
g
al
go
rithm
s
are
div
ided
i
nto
ag
glom
erati
ve
and
div
isi
ve
,
co
rr
e
sp
on
ding
to
th
ei
r
basic
op
era
ti
on
al
natu
re, b
ottom
-
up
or to
p
-
dow
n [6
]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJECE
V
ol.
6
,
No.
3
,
Decem
ber
20
1
7
:
199
–
208
200
Den
sit
y
base
d
cl
us
te
ri
ng
te
chn
i
qu
e
s
a
re
currently
on
the
i
ncr
ease
due
t
o
the
inc
rease
i
n
t
he
requirem
ents
f
or
sp
at
ia
l
data
processi
ng.
Howev
e
r,
m
os
t
of
these
te
c
hn
i
qu
es
are
de
rivati
ves
from
DBSCAN
,
exten
ding
it
accor
ding
to
th
ei
r
operati
ng
do
m
ai
ns
.
E
xtensi
ons
a
re
usu
al
ly
in
te
r
m
s
of
reducti
on
i
n
ti
m
e,
par
am
et
er
auto
m
at
ion
,
in
put
reducti
on
in
te
rm
s
of
featur
e
sel
ect
ion
an
d
abili
ty
to
ha
ndle
va
ried
de
nsi
ti
es.
FD
BSC
AN
[
7]
and
I
DBSC
AN
[
8]
are
tw
o
a
ppr
oach
e
s
con
ce
ntrati
ng
on
the
re
duct
ion
in
tim
e
co
m
po
nen
t.
Both
operate
by
identify
in
g
a
m
ini
m
al
se
t
of
highly
re
pr
ese
ntati
ve
points
f
or
thei
r
op
e
rati
on
rather
tha
n
util
iz
ing
the
e
ntire
data
po
i
nt
s.
In
a
ddit
ion
to
this,
the
I
DBSCA
N
util
i
zed
a
gri
d
bas
ed
data
sel
ect
ion
for
reducin
g
t
he
in
pu
t l
e
vels.
Param
et
er
fine
-
tun
in
g
base
d
al
gorithm
s
include
s
k
-
V
DBS
CAN
[
9],
DB
CLAS
D
[10]
and
AP
SC
AN
[11].
The
k
-
VD
BSC
A
N
is
a
par
am
et
er
fr
ee
te
ch
nique
that
aut
om
at
ic
ally
iden
ti
fies
the
para
m
et
ers.
DBCLAS
D
operates
on
la
rge
data
and
is
a
par
am
et
er
fr
ee
cl
us
te
rin
g
te
chn
i
qu
e
,
ena
bling
gr
a
du
al
cl
us
te
r
exp
a
ns
i
on
on
the
ba
sis
of
the
neig
hbors
an
d
their
de
ns
it
ie
s.
AP
SC
AN
util
iz
es
Affinit
y
Prop
a
gatio
n
te
ch
nique
to ide
ntify t
he c
luster
densi
ti
es b
ase
d o
n
loca
l data va
riat
ions.
A
de
ns
it
y
bas
ed
cl
us
te
ri
ng
t
echn
i
qu
e
ai
ding
in
the
disco
ver
y
of
cl
us
te
r
s
with
var
yi
ng
de
ns
it
ie
s
within
a
si
ng
le
dataset
was
propose
d
by
Zh
u
et
al
in
[12].
I
n
ge
ne
ral,
the
i
nput
data
is
c
onside
red
t
o
c
onta
in
data
distrib
ute
d
in
un
if
orm
den
sit
ie
s.
H
ow
e
ver,
so
m
e
unusual
real
tim
e
data
su
ch
as
po
pu
la
ti
on
m
aps
te
nd
to
con
ta
in
su
c
h
data.
T
his
le
ads
to
a
c
om
pl
ex
issue
,
as
in
creasin
g
the
de
ns
it
y
le
vels
f
or
th
e
entire
process
include
s
severa
l
ou
tl
ie
rs
into
cl
us
te
rs,
w
hile
red
uci
ng
the
den
sit
y
le
vels
m
isses
sever
al
le
gitim
at
e
clu
ste
rs
.
Tw
o
a
ppr
oaches
exist
i
n
li
te
ratur
e
to
s
olv
e
this
is
s
ue,
na
m
el
y
m
od
ify
ing
t
he
al
go
rith
m
app
rop
riat
el
y
an
d
rescali
ng
t
he
da
ta
.
The
te
ch
ni
qu
e
propose
d
in
[12]
util
iz
es
the
la
tt
er
by
rescali
ng
the
da
ta
to
approp
riat
el
y
identify
th
e
cl
us
te
rs
.
DBSC
AN
is
a
pp
li
e
d
to
the
rescale
d
data
t
o
i
den
ti
f
y
cl
us
te
rs.
A
s
i
m
i
la
r
te
ch
nique
t
hat
identifie
s
va
rie
d
de
ns
it
y
base
d
cl
us
te
rs
was
pro
po
se
d
by
L
ouhichi
et
al
.
in
[
13
]
.
T
he
op
erati
on
al
pr
oce
ss
of
this
te
chn
iq
ue
is
div
ide
d
into
two
phase
s.
T
he
first
phase
id
entifi
es
the
de
ns
it
y
le
vels
of
the
input
data
us
in
g
the
expo
nen
ti
a
l
s
pline
te
chn
i
qu
e
on
the
distance
m
a
trix.
S
econd
phase
util
iz
es
the
den
sit
y
values
deter
m
ined
from
the
first
ph
ase
as
loc
al
threshold
pa
ram
et
ers
to
i
den
ti
fy
cl
us
te
r
s.
Othe
r
de
ns
it
y
based
cl
us
te
ri
ng
te
chn
iq
ues
incl
ud
e
VDBSC
A
N
[
14]
, GM
DB
SCAN [
15]
, DDSC
[16], E
D
BSC
AN [
17
]
e
tc
.
This p
ape
r
c
on
centrate
s o
n
de
velo
ping
a d
e
nsi
ty
based
sp
at
i
al
cl
us
te
rin
g
te
chn
i
qu
e
. D
BS
CAN,
bei
ng
the
prec
ur
s
or
of
s
uch
te
ch
niques,
is
ad
opte
d
by
m
os
t
of
the
te
chn
i
ques
in
volvin
g
non
-
un
i
form
cl
us
te
rs.
I
t
was
identifie
d
t
hat
DBSCA
N
has
high
pote
ncy
of
ge
ner
at
in
g
outl
ie
rs.
On
f
ur
th
er
assessm
ent
it
was
ide
ntifie
d
that
sever
al
of
thes
e
ou
tl
ie
rs
eff
ec
ti
vely
fit
into
the
form
ed
cl
us
te
rs.
H
ow
e
ve
r,
they
sti
l
l
re
m
a
in
ou
tl
ie
rs
due
to
the
par
am
et
ers
de
fine
d
f
or
the
cl
us
te
rin
g
pr
oc
ess.
The
pa
r
a
m
et
er
se
tt
ing
process
is
al
ways
optim
al
and
is
identifie
d
by
t
he
data
ex
per
t
s
thr
ough
data
analy
sis
an
d
tr
ia
l
and
e
rror.
As
r
esol
ving
t
his
issue
is
a
c
om
plex
ta
sk
, t
his wo
rk
pr
ese
nts
NRD
BSC
AN w
it
h o
ne
-
cl
ass
S
VM
to r
e
duce the
noise le
vels.
2.
RESEA
R
CH MET
HO
D
Den
sit
y
based
sp
at
ia
l cl
us
te
ri
ng pr
opos
es
a
gro
up
i
ng m
echan
ism
, th
at
o
pe
rates o
n
t
he ba
sis of bot
h
distance a
nd th
e nod
e
d
e
ns
it
y.
Th
e
m
ajo
r
a
dv
antage
of
us
in
g suc
h
a
n
a
ppr
oa
ch
is t
hat it
do
es not rely
on
centr
oid
base
d op
e
rati
ons, he
nc
e the in
co
ns
is
te
ncies in t
he
f
or
m
at
ion
of cl
us
te
rs
are
el
im
i
nated. F
ur
t
her
,
densi
ty
b
ased
c
lusterin
g
is a
n un
s
uper
vise
d
a
ppr
oach with
no pri
or
i
nfor
m
at
ion
requirem
ents.
De
ns
it
y ba
sed
cl
us
te
rin
g
a
ppr
oach
e
s
has
the
abili
ty
to
identi
fy clusters
of a
rb
it
ra
ry sh
a
pes
and sizes
rathe
r
tha
n
bein
g
confine
d
t
o
the
trad
it
io
nal circ
ular
cl
us
te
rs. It
w
as i
de
ntifie
d t
hat D
BSC
AN
base
d
te
ch
niqu
es ex
hib
it
s
high
no
ise
levels.
E
ven th
ough the
b
ase
clusteri
ng
process
w
as
iden
ti
fie
d
to
b
e
eff
ic
ie
nt,
the
noise
levels
ge
ne
rated
by D
BSC
A
N were
foun
d
to
be
e
x
cessi
ve
. T
his
pap
e
r pro
poses
NRDB
SCAN, a
hybri
diz
ed den
sit
y base
d
cl
us
te
rin
g
a
ppr
oach that
i
niti
al
ly
clusters the
data, a
fter
wh
i
ch
it
tries t
o
in
corp
or
at
e
no
ise
into
a
ny
of the
existi
ng d
e
finit
e cluste
rs
, henc
e re
du
ci
ng the
no
ise
levels.
Pr
io
r
to
the
a
c
tu
al
cl
us
te
ri
ng
proces
s,
t
he
i
nput
data
is
processe
d
a
nd
is
co
nv
e
rted
to
t
he
require
d
form
at
.
A
sche
m
a
analy
sis
is
perform
ed
on
t
he
data
to
i
dent
ify
the
data
ty
pes
of
the
c
onte
nts.
T
he
pr
opos
e
d
arch
it
ect
ure
acce
pts
on
ly
nu
m
erical
data,
hen
ce
te
xtu
al
da
ta
are
el
i
m
inate
d
and
cat
eg
ori
cal
and
ordi
na
l
data
are
co
nv
e
rted
to
nu
m
erical
fo
rm
at
s.
The
da
ta
pr
e
-
proces
sing
is
f
ollowe
d
by
the
act
ua
l
cl
us
te
ring
proces
s.
Algorithm
f
or t
he pr
opos
e
d N
RDBSC
A
N
is
sh
ow
n belo
w.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
Den
sit
y B
as
e
d C
lusteri
ng wi
th
In
te
gr
ate
d O
ne
-
Cl
ass
SVM
…
(
K. Nafees
Ah
med
)
20
1
NRDBSCA
N A
lgo
rit
hm
1.
Inp
ut base
data
2.
Inp
ut thres
hold
leve
ls (
minPts,
maxDist)
3.
In
it
iali
ze fi
rst c
luster w
it
h a ra
ndom
node n
4.
neig
hbor
s
identif
yN
ei
ghbo
rs(
n)
5.
If the
neig
hbou
r cou
nt<m
i
nPt
s,
a.
Set n into
a se
para
te
cluster (
Ou
tl
ie
r)
6.
Else
exp
andCl
us
te
r(
n,neig
hb
or
s)
7.
Perf
orm this
process
unti
l all
the
nodes
ar
e
gr
ou
pe
d
int
o a
cl
us
te
r
8.
For
e
ac
h
ide
nt
if
ie
d
cl
us
te
r C
a.
If d
e
ns
it
y o
f
C
< 1
i.
no
ise
C
9.
Un
ti
l N
oise is
no
t e
m
pty
a.
For
e
ac
h
ide
nt
if
ie
d
cl
us
te
r C
i.
Predict
ion
pre
dict(
C,no
ise
)
ii.
Ad
d all
true
pre
dicti
on
s
to
cl
ust
er C
iii.
Delet
e corr
e
spondin
g
e
ntries
from
noise
b.
if
n
o
e
ntries
are
d
el
et
ed
for
t
he
last tw
o
it
er
at
ion
s
i.
All
oca
te
new c
lusters f
or eac
h
e
ntry in
nois
e
ii.
Em
pty noise
fu
ncti
on i
de
nti
fyN
ei
ghbo
r
s (n)
1.
For all
no
des n
1
in
data
a.
If d
ist
ance(
n,n
1)
<m
ax
Dist
i.
Ad
d n1 to
neig
hbor
List
2.
ret
ur
n nei
ghbo
rList
fu
ncti
on
e
xpandCl
uster(
n,ne
ighbors
)
1.
Ad
d n to t
he
cl
us
te
r C
2.
For
e
ac
h neig
hbour
n1
a.
neig
hborL1
identif
yN
ei
ghbors(
n1)
b.
if
n
ei
ghborL
1 count
>=
minP
ts
i.
Ad
d n1 to C
fu
ncti
on
pr
edi
ct
(C,no
ise
)
1.
In
it
iali
ze one
-
c
lass
SVM
wi
th
po
ly
nomial k
er
nel
2.
Set the
de
gr
ee
of ker
nel to be
the d
i
mensi
ons
o
f C
3.
Tra
in
SVM w
it
h
C
4.
Predict
ions
Ap
ply noise t
o t
ra
ine
d
ker
nel
5.
Ret
ur
n Pre
dicti
on
s
2.1.
I
nitial Le
vel C
lu
ster
Fo
rmul
at
i
on
usi
ng
DBS
CAN
The
pr
e
-
pr
oce
ssed
data
is
a
naly
zed
a
nd
the
m
axi
m
u
m
acce
ptable
dis
ta
nce
(
ma
x
Dis
t
)
an
d
the
m
ini
m
u
m
neig
hbor
requirem
ents
(
mi
nPts
)
are
identifie
d
us
in
g
the
data
distribu
ti
on.
Howe
ver,
accuratel
y
identify
in
g
the
best
par
am
et
ers
is
no
t
po
s
sib
le
in
a
sing
le
it
erati
on.
This
is
a
tria
l
and
error
base
d
fine
-
tun
i
ng
process
.
F
ur
t
he
r,
the
se
par
am
et
ers
va
ry
with
t
he
dataset
bei
ng
us
ed
.
Hen
ce
distrib
ution
ba
sed
tra
ns
ie
nt
va
lues
are
init
ia
ll
y
identifie
d
an
d
th
e
final
par
am
e
te
rs
are
identif
ie
d
by
m
et
ho
di
cal
ly
increasing
a
nd
decr
eas
ing
th
e
par
am
et
er v
al
ue
s to
i
den
ti
fy t
he best
par
am
et
er s
et
for t
he c
urren
t
data
unde
r
a
naly
sis.
2.1.1.
Clus
ter
Ident
ific
at
i
on
Pha
se
In
t
his
phase
,
t
he
process
be
gi
ns
with
a
ra
nd
om
no
de
i
n
the
dataset
.
T
he
i
niti
al
cl
us
te
r
is
com
po
sed
of
t
his
sin
gle
node
n
.
Nei
ghbors
of
the
sel
ect
ed
no
de
(
n
)
a
re
ide
ntifie
d
usi
ng
m
axD
ist
as
the
th
res
ho
l
d.
If
t
he
node
sat
isfie
s
the
m
ini
m
u
m
neighbor
requir
e
m
ents
(
mi
nPt
s
),
it
is
c
onside
red
to
be
a
pa
r
t
of
a
cl
us
te
r
a
nd
not
an ou
tl
ie
r. He
nc
e, this is
f
ollo
wed b
y t
he
cl
ust
er expa
ns
io
n ph
a
se.
2.1.2.
Clus
ter
Expansio
n Phase
This
ph
ase
fin
ds
the
nei
ghbors
of
eac
h
no
de
in
n.
If
e
ach
of
t
hese
nodes
sat
is
fies
the
mi
nPt
s
requirem
ent
they
are
inc
orporated
as
a
n
entit
y
in
the
cl
us
te
r.
T
his
proces
s
is
c
on
ti
nue
d
un
ti
l
al
l
the
appr
opriat
e
nodes
are
gro
up
e
d
into
t
he
cl
ust
er.
H
ow
e
ve
r,
sever
al
cl
ust
er
s
m
igh
t
exist
in
a
dataset
.
H
ence
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJECE
V
ol.
6
,
No.
3
,
Decem
ber
20
1
7
:
199
–
208
202
so
m
e
po
ints
w
ou
l
d
rem
ai
n
ou
tsi
de
t
he
c
ompo
s
ed
cl
us
te
r
.
A
rand
om
su
ch
po
i
nt
is
c
ons
idere
d
as
the
ba
se
f
or
the
ne
xt
cl
us
te
r.
T
he
nei
ghbo
r
ide
ntific
at
ion
and
cl
us
te
r
e
xpan
sio
n
phase
are
re
peated
t
o
identify
the
points
corres
pondin
g t
o
the
ne
w
cl
ust
er.
T
his
proce
ss is r
e
peated
unti
l al
l t
he
poin
ts are cat
e
gorized int
o
a cl
us
te
r.
Th
ough
al
l
points
are
cat
eg
or
iz
e
d
into
cl
us
te
rs
,
not
al
l
cl
us
te
rs
w
ould
be
c
om
po
se
d
with
high
or
eve
n
consi
der
a
ble
de
ns
it
ie
s.
So
m
e
cl
us
te
rs
w
ou
l
d
be
c
om
po
sed
of
sing
le
node.
T
hese
no
de
s
are
inten
de
d
to
be
ou
tl
ie
rs.
Ma
jor
reasons
f
or
th
e
occ
urren
ce
of
outl
ie
rs
are
i
m
pr
op
er
pa
ra
m
et
er
tun
in
g,
t
heir
act
ual
pr
e
sence
or
inap
pro
pr
ia
te
node
validat
io
n
in
the
e
xpan
sion
ph
a
se.
T
houg
h
the
act
ua
l
pr
ese
nce
of
ou
tl
ie
rs
needs
to
be
consi
der
e
d,
th
e
oth
e
r
tw
o
is
su
es
al
s
o
play
a
vital
r
ole
in
the
occ
urre
nc
e
of
outl
ie
rs
i
n
DB
SCA
N.
Hen
ce
a
seco
nd
le
vel
e
xam
inati
on
w
ould
i
ncor
porate
seve
ral
no
de
s
that
c
ou
l
d
ha
ve
been
com
po
ne
nts
of
the
e
xisti
ng
cl
us
te
rs, le
adin
g
to
a
reducti
on in
the
outl
ie
rs.
2.2.
On
e
-
Cl
as
s
SVM
ba
se
d
N
oise
Redu
c
tio
n
DBSCA
N
el
im
inate
s
severa
l
no
des
,
co
nsi
der
in
g
them
as
ou
tl
ie
rs.
Howe
ver,
they
m
igh
t
no
t
necessa
rily
correspo
nd
to
out
li
ers,
rat
her
th
ey
co
uld
be
c
om
po
ne
nts
of
the
def
ine
d
cl
us
te
rs
.
Util
iz
ing
th
e
conve
ntion
al
c
onditi
on
al
che
cks
util
iz
ed
by
DBSCA
N
a
gain
ulti
m
a
te
l
y
has
t
he
sam
e
eff
ect
s
.
Hence
th
e
pro
po
se
d
NR
DBSCA
N
inc
orp
or
at
es
a
m
achine
le
ar
ning
cl
assifi
er,
On
e
-
Cl
ass
Suppo
rt
Vector
Ma
chine
(S
VM
) for th
e
cat
egorizat
ion
process
.
Util
iz
ing
all
the clusters
for
tr
ai
nin
g a bi
nar
y
/
m
ulti
-
cl
ass class
ifie
r
an
d
t
hen p
er
form
ing
classi
ficat
io
n
on
t
he
detect
e
d
ou
tl
ie
rs
has
se
ver
al
dow
ns
ide
s.
The
m
ajor
issue
is
that
bi
na
ry/m
ulti
-
cl
ass
cl
assifi
er
de
fin
it
el
y
cat
egorizes
the
data
i
nto
a
ny
one
of
the
gr
oups
.
Hen
ce
a
ll
the
outl
ie
rs
would
de
finite
ly
be
groupe
d
into
a
cl
us
te
r,
le
a
ding
to
z
er
o
ou
tl
ie
rs.
H
ow
e
ver,
the
pro
pose
d
work
is
inte
nded
on
gro
upin
g
on
ly
the
a
pp
ropr
ia
te
nodes
a
nd
retai
nin
g
the
outl
ie
rs.
He
nce
a
one
-
cl
ass
cl
assif
ie
r
w
ou
l
d
be
t
he
best
s
uited
appr
oach.
O
ne
-
cl
ass
cl
assifi
er
is
a
s
pec
ia
l
case
of
cl
assifi
ers,
w
he
re
the
patte
r
n
of
a
si
ng
le
cl
a
ss
w
ould
be
w
el
l
known,
w
hi
le
the
patte
rn
s
that
do
no
t c
onfi
ne
t
o
the
w
el
l
-
known
traine
d patt
ern
s
wo
uld
be c
onsidere
d
a
s
ou
tl
ie
rs.
Ou
t
pu
t
f
r
om
D
BSC
AN
is
in
t
erm
s
of
cl
us
te
r
s.
Data
withi
n
cl
us
te
rs
h
a
ve
obvi
ous
associat
ion
s
,
he
nce
these
can
be
use
d
as
the
trai
ning
data
f
or
the
cl
assifi
ers.
The
noise
re
duct
ion
phase
op
erates
by
trai
nin
g
th
e
one
-
cl
ass
SVM
us
in
g
data
from
on
e
cl
us
t
er
a
nd
perfor
m
ing
pr
e
dicti
ons
on
the
detect
ed
ou
tl
ie
rs.
Ou
tl
ie
rs
cat
egoriz
ed
as
com
po
nen
ts
of
the
cl
us
te
r
us
e
d
for
trai
ni
ng
a
re
inco
rpo
rated
into
the
cl
us
te
r.
T
he
re
sidu
al
ou
tl
ie
rs
are
c
onside
red
for
processin
g
by
tr
ai
nin
g
the
on
e
-
cl
ass
SV
M
w
it
h
data
fr
om
t
he
ne
xt
cl
us
te
r
.
This
process
is co
nt
inu
e
d for all
th
e d
e
fine
d
cl
us
t
ers w
it
h
c
onsid
erab
le
de
ns
it
y l
evels.
One
-
cl
ass
SVM
was
sug
gest
ed
by
Sc
holk
opf
et
al
.
i
n
[18
]
.
It
op
e
rates
by
sepa
rati
ng
th
e
data
points
from
the
or
igin
and
c
onside
ring
t
he
ori
gi
n
a
lon
e
as
t
he
sec
ond
cl
ass.
The
base
op
e
rati
onal
natur
e
of
on
e
-
cl
ass
SV
M
is
to
m
a
xim
iz
e
the
distance
f
ro
m
the
hype
rp
la
ne
cre
at
ed
by
the
dat
a
po
i
nts
to
the
or
i
gin
.
T
he
res
ultant
of
this
is
a
bi
nar
y
functi
on
that
effe
ct
ively
captu
res
t
he
re
gions
in
th
e
input
s
pace,
retu
rn
i
ng
+
1
for
data
con
ta
ine
d
in
t
he
hype
rp
la
ne
an
d
-
1
f
or
ot
her
s.
O
ne
-
cl
ass
S
VM
use
d
in
t
he
NRD
BSC
AN
util
iz
es
th
e
po
ly
nom
ia
l ker
nel [1
9], as
t
he
input
data
has t
he
te
nden
cy
to
contai
n
se
vera
l dim
ension
s.
No
t
al
l
data
a
r
e
ex
pected
to
be
c
om
po
ne
nts
of
the
cl
us
te
rs
.
A
fter
the
e
ntire
process,
s
om
e
resid
ual
da
ta
rem
ai
ns
, w
hic
h a
re cate
gorize
d
as
outl
ie
rs.
3.
RESU
LT
S
A
ND AN
ALYSIS
DBSCA
N
was
i
m
ple
m
ented
in
C
#.
N
ET
a
nd
the
no
ise
re
duct
ion
c
om
po
ne
nts
w
ere
i
ncor
porated
int
o
the
op
e
rati
ona
l
pr
oces
s.
A
na
ly
sis
of
the
N
RDBSC
A
N
w
as
perform
ed
by
us
in
g
f
our
ben
c
hm
ark
dat
aset
s.
Cl
us
te
rin
g
sp
e
ci
fic
dataset
s
su
ch
as
I
ris
and
Ba
nan
a
an
d
sp
at
ia
l
dataset
s
su
c
h
as
Qu
a
ke
and
F
or
e
st
were
us
e
d
for
ide
ntifyi
ng
the ef
fici
ency
of the al
gorith
m
.
Cl
us
te
r
de
ns
it
ie
s
and
t
he
nu
m
ber
of
cl
us
te
rs
obta
ine
d
f
rom
each
of
the
dataset
s
us
i
ng
NR
D
BSC
A
N
is
sh
own
in
fi
gures
(1
-
4).
Cl
us
te
r
de
ns
it
ie
s
cor
re
spo
nd
to
the
nu
m
ber
of
nodes
in
eac
h
of
th
e
form
ulate
d
cl
us
te
r.
A
de
nsi
ty
of
one
in
dicat
es
an
ou
tl
ie
r.
It
co
uld
be
ob
se
r
ved
t
ha
t
the
cl
us
te
rs
f
or
m
ulate
d
us
in
g
the
pro
po
se
d
a
ppr
oach
ex
hib
it
s
low
outl
ie
r
l
evels
a
nd
cl
ust
ers
of
c
onsi
der
a
ble
den
sit
ie
s.
I
ris
an
d
Qu
a
ke,
con
ta
ini
ng
lo
w
va
riat
ion
s
e
xh
ibit
low
outl
ie
r
le
vels,
w
hile
Fo
r
est
an
d
Ba
nan
a
ex
hib
it
s
high
var
ia
ti
ons
,
he
nc
e
high
ou
tl
ie
rs.
Howe
ver, it
could
be o
bs
er
ve
d
that t
he
cl
us
te
rs othe
r
tha
n ou
tl
ie
rs
e
xhibit
h
ig
h de
ns
it
y l
evels.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
Den
sit
y B
as
e
d C
lusteri
ng wi
th
In
te
gr
ate
d O
ne
-
Cl
ass
SVM
…
(
K. Nafees
Ah
med
)
203
Figure
1. Cl
us
t
er
Den
sit
y (
Ir
is
)
Figure
2. Cl
us
t
er
Den
sit
y (Ba
nan
a
)
Figure
3. Cl
us
t
er
Den
sit
y (
Q
ua
ke)
50
54
11
23
4
7
1
0
10
20
30
40
50
60
1
2
3
4
5
6
7
No.
of Nod
es
Cluste
r
Number
Cl
us
t
er
Dens
ity
(Iris)
5066
131
9
40
16
13
13
6
1
1
1
1
1
1
0
1000
2000
3000
4000
5000
6000
1
2
3
4
5
6
7
8
9
10
11
12
13
14
No.
of Nod
es
Cluste
r
Number
Cl
us
t
er
Dens
ity
(Ba
nana)
2045
21
8
7
9
18
12
22
26
9
1
0
500
1000
1500
2000
2500
1
2
3
4
5
6
7
8
9
10
11
No.
of Nod
es
Cluste
r
Number
Cl
us
t
er
Dens
ity
(Quak
e)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJECE
V
ol.
6
,
No.
3
,
Decem
ber
20
1
7
:
199
–
208
204
Figure
4. Cl
us
t
er
Den
sit
y (
For
est
)
In
tra
-
cl
us
te
r
ra
diu
s
of
the
pro
po
s
ed
NRDBS
CAN
is
presen
te
d
in
fig
ur
es
(
5
-
8).
It
co
uld
be
obser
ve
d
that
the
propos
ed
te
ch
nique
exh
i
bits
low
to
m
od
erate
intra
cl
us
te
r
distanc
e
le
vels.
Mod
e
rate
le
vels
are
du
e
t
o
the
var
ie
d
s
ha
ped
cl
us
te
rs
f
or
m
ed
by
the
al
go
rithm
.
Int
er
-
cl
us
te
r
dist
ance
ex
hib
it
ed
by
the
al
gori
thm
is
pr
ese
nted
in
fi
gure
9.
It
c
ould
be
obser
ve
d
that
lo
w
i
nter
cl
us
te
r
dista
nc
es
are
ex
hib
it
ed
by
NRDBS
CAN.
This
validat
es
our
cl
ai
m
o
f va
ried
s
ha
ped clu
ste
rs
wit
h diff
e
ren
t
de
ns
it
y l
evels.
Figure
5. I
ntra C
luster Radi
us (Iris)
494
4
5
4
1
1
1
1
1
1
1
1
1
1
0
100
200
300
400
500
600
1
2
3
4
5
6
7
8
9
10
11
12
13
14
No.
of Nod
es
Cluste
r
Number
Cl
us
t
er
Dens
ity
(F
or
es
t)
0
0.5
1
1.5
2
2.5
3
1
2
3
4
5
6
7
Node Di
stance
Cluste
r
Number
In
tr
aClu
s
t
er
Ra
di
us
(Iris)
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
Den
sit
y B
as
e
d C
lusteri
ng wi
th
In
te
gr
ate
d O
ne
-
Cl
ass
SVM
…
(
K. Nafees
Ah
med
)
205
Figure
6. I
ntra C
luster Radi
us (Bana
na
)
Figure
7. I
ntra C
luster
Ra
di
us (Qua
ke)
Figure
8. I
ntra C
luster Radi
us (F
or
est
)
0
0.5
1
1.5
2
2.5
3
3.5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Node Di
stance
Cluste
r
Number
In
tr
aClu
s
t
er
Ra
di
us
(Ba
nana)
0
0.5
1
1.5
2
2.5
3
3.5
4
1
2
3
4
5
6
7
8
9
10
11
Node Di
stance
Cluste
r
Number
In
tr
aClu
s
t
er
Ra
di
us
(Quak
e)
0
100
200
300
400
500
600
1
2
3
4
5
6
7
8
9
10
11
12
13
14
Node Di
stance
Cluste
r
Number
In
tr
aClu
s
t
er
Ra
di
us
(F
or
es
t)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJECE
V
ol.
6
,
No.
3
,
Decem
ber
20
1
7
:
199
–
208
206
Figure
9. I
nter C
luster
Dista
nc
e
A
c
om
par
ison
in
te
rm
s
of
ti
m
e
is
carried
ou
t
betwee
n
D
BSC
AN,
Mo
di
fied
PS
O
base
d
DBSCA
N
[20]
an
d
the
pro
posed
NR
D
BSC
AN
is
pr
e
sented
i
n
fi
gur
e
10.
It
c
ou
l
d
be
obse
rv
e
d
t
ha
t
the
tim
e
ta
k
en
f
or
m
od
ifie
d
PSO
is
the
hig
hest
.
Howe
ver,
th
e
tim
e
ta
ken
by
NRDBSCA
N
is
higher
c
om
par
ed
to
DB
SCAN.
Eve
n
-
th
ough
t
hat
is
the
ca
se,
the
diff
e
ren
c
e
is
in
te
rm
s
of
0.3
sec
(m
axim
u
m
).
Hen
ce
t
he
sig
nifica
nce
of
the
tim
e increase is co
ns
ide
re
d
to
b
e l
ow.
Figure
10. Tim
e Com
par
iso
n
A
c
om
par
ison i
n
te
rm
s
of
th
e n
um
ber
o
f
cl
ust
ers
ge
ne
rated b
y
DBSC
A
N
a
nd
NR
DBSCA
N
is
s
how
n
in
fi
gure
11.
It
co
uld
be
obse
rv
e
d
t
hat
the
num
ber
of
cl
ust
ers
gen
e
rated
by
NRDBSC
A
N
is
at
-
le
ast
60%
le
s
s
than
the
c
onve
ntion
al
DBSC
AN.
This
e
xhibit
s
the
ef
fecti
ve
no
ise
re
duct
ion
le
vels
e
xhibit
ed
by
the
pr
opose
d
appr
oach.
0
0.5
1
1.5
2
2.5
3
3.5
4
Iris
Ba
nan
a
Qua
ke
Forest
Distance
In
t
er
Cl
us
t
er
Di
s
t
ance
Iris
Ba
nan
a
Qua
ke
Forest
Ti
m
e (D
B
SCAN
)
0.002
0.046
0.0361
0.0126
Ti
m
e
(
N
R
DBSCA
N)
0.005
0.3605
0.0706
0.0217
Ti
m
e (M
odifi
ed
PSO)
0.193
3.972
3.421
1.794
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
Ti
me
(
sec)
Evaluation Warning : The document was created with Spire.PDF for Python.
IJ
-
ICT
IS
S
N:
22
52
-
8776
Den
sit
y B
as
e
d C
lusteri
ng wi
th
In
te
gr
ate
d O
ne
-
Cl
ass
SVM
…
(
K. Nafees
Ah
med
)
207
Figure
11. C
om
par
ison
in
ter
m
s o
f
num
ber
of clusters
4.
CONCL
US
I
O
N
Cl
us
te
rin
g,
bei
ng
one
of
the
m
ajo
r
te
ch
niqu
es
f
or
data
ana
ly
sis
has
see
n
sever
al
ada
ptati
on
s
due
t
o
the
cha
ng
i
ng
const
raints.
T
he
m
ajo
r
a
da
ptati
on
s
i
nclu
de
ide
ntifyi
ng
var
ie
d
sh
a
pe
d
cl
us
te
r
s.
T
hi
s
pap
e
r
pro
po
ses
a
n
en
han
cem
ent
of
the
DBSCA
N
that
is
a
pio
neer
al
go
rithm
in
i
den
ti
fyi
ng
va
ried
sh
a
ped
a
nd
v
arie
d
densi
ty
cl
us
te
rs.
The
m
ajo
r
dow
ns
id
e
of
DB
SCAN
was
ob
serv
e
d
to
be
it
s
requirem
ent
fo
r
acc
ur
at
e
pa
r
a
m
et
er
set
ti
ng
.
A
sli
ghtl
y
var
ie
d
set
ti
ng
res
ults
in
th
e
increa
se
in
outl
ie
rs.
Howe
ve
r,
i
den
ti
fyi
ng
the
pe
rfec
t
pa
r
a
m
et
er
is
no
t
feasible
.
Hen
ce
the
propose
d
NR
DB
SCAN
e
nhanc
es
the
DBSCA
N
ap
proac
h
by
introdu
ci
ng
a
no
ise
reducti
on
c
ompone
nt
ba
sed
on
on
e
-
cl
ass
cl
a
ssifie
r
m
od
el
.
On
e
-
cl
ass
SVM
was
us
e
d
to
pe
rfor
m
this
proces
s
.
Re
su
lt
s
ex
hib
i
ts
sign
ific
a
nt
r
edu
ct
io
n
i
n
th
e
no
ise
le
vels.
Current
al
go
r
it
h
m
op
erates
eff
ect
ively
on
fixe
d
densi
ty
cl
us
te
r
s.
F
uture
wor
ks
will
co
nce
ntr
at
e
on
po
rting
t
he
al
gorithm
to
op
e
rate
on
da
ta
set
s
with
va
ried
densi
ti
es and e
ff
ect
ively
ide
nt
ify
clusters
with
var
ie
d den
sit
y l
evels.
REFERE
NCE
S
[1]
J
.
A
.
Hart
iga
n
an
d
M.A.
W
ong
,
“
A
lgori
thm AS
1
36:
A k
-
m
ea
ns
C
luste
ring
A
lgorit
hm
,
”
Journal
of
Roy
al
Sta
ti
sti
ca
l
Soci
e
ty
,
S
erie
s C
(
Appl
ie
d
Sta
ti
sti
cs)
vol.
2
8,
pp.
1
00
-
108
,
Jan
197
9
.
[2]
C.
P.
W
ei
,
et
al.,
“
Empiric
al
Com
par
ison
of
Fas
t
Cluste
ring
Algorit
hm
s
for
La
rge
dat
a
sets,
”
in
S
y
stem
Scie
nce
s
,
2000
IEEE Proc
ee
ding
of
33
rd
A
nnual
Hawaii
In
te
rnational
Conf
ere
nc
e
,
2000
,
pp
.
1
-
1
0.
[3]
M.E
ster,
e
t
al.,
“
A
Densit
y
B
as
ed
Algorit
hm
for
Discove
ring
Cluste
rs
in
La
rge
Spati
al
Dat
aba
se
s
with
Noise,
”
in
2
nd
Inte
rnat
ional Confe
ren
ce on Knowledge Disc
ove
ry
and
Data
Mini
ng.
KDD
-
1996
,
vol
.
96
,
pp
.
226
-
231.
1996
.
[4]
A.
Hinnebur
g
an
d
D.A.
Ke
im,
“
An
Eff
icient
Approac
h
to
Clust
e
ring
in
La
rg
e
M
ult
imedia
Dat
ab
ase
s
with
Noise
,
”
in
Int
ernati
onal
Confe
renc
e
on
K
nowle
gde
Disco
ve
ry
and
Data
M
ini
ng,
KDD
-
1998
,
vol
.
98
,
pp
.
58
-
65,
1998
.
[5]
M.
Ankerst,
et
al
.
,
“
OP
TICS:
Ord
eri
ng
Points
to
Ide
nti
f
y
th
e
Clust
eri
ng
Struct
ur
e,”
in
ACM
SIGMO
D R
ec
ord 1999,
vol.
28
,
pp
.
49
-
6
0,
1999
.
[6]
E.
Güngör
and
A.
Özm
en
,
“
Distance
and
Densi
t
y
b
ase
d
Clust
er
ing
Algorit
hm
u
sing
Gauss
ia
n
Kerne
l,”
E
xpe
rt
Syste
ms
wit
h
Ap
pil
cations,
vo
l. 1
,
pp
.
10
-
20
,
201
7.
[7]
S.
Zhou,
e
t
a
l.,
“
FD
BS
C
AN
:
A Fast
DBS
CAN
Algorit
hm
,
”
Ru
an Jian
Xue
B
ao
,
v
ol.
11
,
pp
.
735
-
7
44,
2000
.
[8]
C.
F.
Tsai
n
and
H.F.
Yeh,
“
Npus
t:
An
Eff
icient
Cluste
ring
Algo
rit
hm
using
Part
it
ion
Spa
ce
T
echnique
for
L
arg
e
Data
base
s,
”
in
I
nte
rnational
Co
nfe
renc
e
on
Indu
strial,
Engi
n
eer
ing
and
Other
Appl
ic
a
ti
ons
of
Appl
ie
d
In
te
l
li
ge
nt
Syste
ms
,
pp
.
787
-
796,
2009
,
Spri
nger
Ber
li
n
Heid
el
ber
g
.
[9]
A.R.
Chowdhury
,
et
al.,
“
An
Eff
ic
i
ent
Met
hod
forSs
ubje
ct
ive
l
y
Choosing
Para
m
et
er
„k‟Autom
at
ic
a
lly
i
n
VD
BS
CA
N
(
Vari
ed
Densit
y
B
a
sed
Spa
ti
al
Cluste
ring
of
Applicati
ons
with
Noise)Algori
thm,”
in
Computer
and
Aut
omation
Enginee
ring,
2010,
I
CCAE
2010
.
S
econd Int
ernat
iona
l
Conf
ere
nce, IE
EE
,
pp.
38
-
41.
[10]
M.
Parimala,
et
al.,
“
A
S
urve
y
on
D
ensity
B
a
sed
C
luste
ring
A
lgori
thms
for
M
ini
ng
L
arg
e
S
pat
i
al
Da
ta
base
s
,
”
Inte
rnational
Jo
urnal
of
Adv
an
c
ed
Sc
ie
n
ce
and
T
ec
hnolog
y,
vol.
31,
pp
.
59
-
66
,
2
011.
[11]
X.
Chen,
et
al
.
,
“
APSCA
N:
A
P
ara
m
et
er
F
re
e
A
lgori
thm
for
C
luste
ring
,”
Pattern
Re
cognition
Let
te
rs
,
vol.
32,
pp
.
973
-
986,
2011
.
[12]
Y.
Zhu,
et
al.,
“
Density
-
r
atio
B
ase
d
C
luste
ri
ng
for
Discove
ring
C
luste
rs
with
V
ar
y
ing
D
ensit
ie
s
,”
Pat
t
ern
Re
cogn
it
ion
,
vo
l
.
60
,
pp
.
983
-
99
7,
2016
.
[13]
S.
Louhi
ch
i,
e
t
a
l.
,
“
Uns
uper
vise
d
Vari
ed
Densit
y
B
ase
d
C
luste
ri
ng
A
lgori
thm
using
S
pli
ne
,”
Pa
ttern
Re
cogn
it
ion
Lett
ers
,
2016
.
0
50
100
150
200
250
300
Iris
Ba
nan
a
Qua
ke
Forest
Cluste
r
Number
No of Clus
t
er
s
# C
l
us
te
r
s
(
DBSCA
N)
# C
l
us
te
r
s
(
NRD
BS
CAN)
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
IJECE
V
ol.
6
,
No.
3
,
Decem
ber
20
1
7
:
199
–
208
208
[14]
P.
Li
u,
e
t
al.,
“
VD
BS
CA
N:
V
a
rie
d
D
ensity
B
a
sed
S
pat
ia
l
C
lu
steri
ng
of
A
pplications
with
N
oise
,”
i
n
S
erv
ice
Syste
ms
and
S
erv
ice
Manag
eme
n
t
,
2007
,
I
EEE
Int
ernati
onal
Confer
enc
e,
2007
,
pp.
1
-
4.
[15]
C.
Xiao
y
um
,
et
al.
,
“
GM
DBS
C
AN
:
M
ult
i
-
Densit
y
DBS
CAN
C
luste
r
B
ase
d
on
Grid,
”
in
e
-
Busi
ness
Engi
nee
rin
g
,
2008,
ICE
BE 20
08.
IE
EE Int
ern
ati
onal
Con
fe
ren
ce
,
2008,
pp.
78
0
-
783.
[16]
B.
Borah
and
D.K.
Bhatta
cha
r
yy
a
,
“
DD
SC
:
A
D
ensity
D
i
ffe
r
e
nti
ated
S
patial
C
luste
ring
Te
ch
nique
,
”
Journal
of
compute
rs
,
vol
.
3,
pp
.
72
-
79
,
20
08.
[17]
A.
Ram,
e
t
al
.
,
“
An
Enha
nce
d
D
ensity
B
ase
d
S
pat
ial
C
lust
er
ing
of
A
pplica
t
ions
with
N
oise
,”
in
Ad
vanced
Computing
Conf
ere
nce,
2009
.
I
A
CC 2009.
I
EEE
Inte
rnational
,
20
09,
pp
.
1475
-
14
78.
[18]
B.
Schölkopf
,
et
al.
,
“
Esti
m
at
ing
the
S
upport
of
a
H
igh
-
dimensional
D
istri
buti
on
,
”
Neural
Computati
on
,
2001
,
pp.
1443
-
1471.
[19]
L.
M.
Man
evi
t
z
and
M.
Yous
ef,
“
One
-
cl
ass
SV
Ms
for
D
ocument
Cla
ss
ifica
t
ion
,
”
Journal
of
M
achi
ne
Learning
Re
search
,
2001,
vol.
2
,
pp
.
139
-
1
54.
[20]
K.
Nafees
Ahm
ed
and
T.
Abdul
Raz
ak
,
“
Densit
y
Based
Clust
eri
n
g
using
Modifi
e
d
PS
O
base
d
Ne
ighbor
Sel
ec
t
ion
,
”
Inte
rnational
Jo
urnal
of
Comput
er
Scienc
e
and
E
ngine
ering
,
vol
.
9,
pp
.
192
-
199
.
Evaluation Warning : The document was created with Spire.PDF for Python.