T
E
L
KO
M
NI
K
A
,
V
o
l.
1
4
,
N
o.
3
,
S
ept
em
ber
20
1
6
,
pp.
11
50
~
115
6
I
S
S
N
:
1
693
-
6
930
,
ac
c
r
edi
t
ed
A
b
y
D
IK
T
I,
D
e
c
r
e
e
N
o
:
58/
D
I
K
T
I
/
K
ep/
2013
D
O
I
:
10.
12928/
T
E
LK
O
M
N
I
K
A
.
v
1
4
i
3
.
3805
11
50
R
ec
ei
v
ed
M
ar
c
h
1
4
,
201
6
;
R
ev
i
s
ed
May
3
1
,
2
01
6
;
A
c
c
ept
e
d
J
une
1
5
,
20
1
6
A
N
e
w
S
e
m
i
-
su
p
er
v
ised
Clu
s
t
e
r
in
g
A
l
g
o
r
it
h
m
Based
o
n
V
ar
iat
io
n
al Ba
y
esian
a
n
d
I
t
s
A
p
p
licat
i
o
n
S
h
o
u
l
i
n
Y
i
n
1
,
Ji
e L
i
u
*
2
,
L
i
n
T
e
n
g
3
S
of
t
w
ar
e C
ol
l
e
ge
,
S
h
eny
ang N
or
m
al
U
n
i
ve
r
si
t
y
,
N
o.
253,
H
uangH
e
B
ei
S
t
r
eet
,
H
uangG
u D
i
s
t
r
i
c
t
,
S
heny
ang
,
P
.
C
1100
34
-
C
hi
n
a
*
C
or
r
es
pond
i
ng
aut
hor
, e
-
m
ai
l
:
nan1
27@
s
ohu.
c
o
m
*
1
,
3
5272
0214@
q
q.
c
om
2
,
153
255
4069
@
qq.
c
o
m
3
A
b
st
r
act
B
ic
lu
s
t
e
r
in
g
a
l
gor
i
t
hm
i
s
pr
o
pos
e
d f
or
di
s
c
ov
er
i
ng m
at
r
i
x
w
i
t
h bi
o
l
ogi
c
a
l
s
i
gni
f
i
c
a
nc
e
i
n gene
ex
pr
e
s
s
i
on
d
at
a
m
at
r
i
x
an
d
i
t
i
s
us
e
d
w
i
de
l
y
i
n
m
ac
hi
ne
l
ear
ni
ng
w
h
i
c
h
c
an
c
l
u
s
t
er
t
he
r
ow
and
c
o
l
um
n
o
f
m
at
r
i
x
.
I
n
or
der
t
o
f
ur
t
her
i
m
pr
ov
e
t
he
per
f
or
m
anc
e o
f
b
i
c
l
us
t
er
i
ng
al
gor
i
t
hm
,
t
h
i
s
pa
per
pr
opos
es
a
s
em
i
-
s
uper
v
i
s
ed
c
l
us
t
er
i
ng
al
gor
i
t
hm
bas
e
d on
v
ar
i
at
i
ona
l
B
ay
e
s
i
an
.
F
i
r
s
t
l
y
,
i
t
i
nt
r
od
uc
e
s
s
u
ppl
em
ent
a
r
y
i
nf
or
m
at
i
on
of
r
ow
an
d
c
ol
u
m
n
f
or
b
i
c
l
us
t
er
i
n
g
pr
oc
e
s
s
a
nd
r
e
pr
es
ent
s
c
or
r
e
s
pon
di
n
g
j
oi
nt
di
s
t
r
i
but
i
o
n
pr
obab
i
l
i
t
y
m
od
el
.
I
n
a
ddi
t
i
on,
i
t
e
s
t
i
m
at
e
s
t
he
p
ar
am
et
er
of
j
oi
n
t
di
s
t
r
i
but
i
on
pr
oba
bi
l
i
t
y
m
odel
ba
s
e
d
o
n
v
ar
i
a
t
i
on
al
B
ay
e
s
i
a
n
l
ear
ni
ng
m
et
hod.
F
i
na
l
l
y
,
i
t
es
t
i
m
at
es
t
he
per
f
or
m
anc
e
of
pr
op
os
e
d
a
l
gor
i
t
hm
t
hr
ou
g
h
s
y
nt
he
s
i
z
ed
dat
a
an
d r
eal
ge
n
e ex
pr
e
s
s
i
on da
t
a s
et
.
E
x
per
i
m
ent
s
s
how
t
h
at
n
or
m
al
i
z
ed
m
ut
ual
i
nf
or
m
at
i
on
o
f
t
hi
s
p
aper
’
s
new
m
et
ho
d i
s
bet
t
er
t
han
r
el
e
v
a
nt
bi
c
l
us
t
er
i
n
g al
g
or
i
t
hm
s
f
or
bi
c
l
us
t
er
i
ng a
n
al
y
s
i
s
.
K
eyw
o
r
d
s:
b
i
c
l
us
t
er
i
n
g al
g
or
i
t
hm
,
v
ar
i
a
t
i
onal
b
ay
es
ia
n
,
j
o
i
nt
d
i
s
t
r
i
but
i
on pr
o
bab
i
l
i
t
y
,
s
e
mi
-
s
up
e
r
v
is
ed
c
l
us
t
e
r
in
g
C
o
p
y
r
i
g
h
t
©
20
16 U
n
i
ver
si
t
a
s A
h
mad
D
ah
l
an
.
A
l
l
r
i
g
h
t
s r
eser
ved
.
1.
I
n
tr
o
d
u
c
ti
o
n
I
n t
h
e
gi
v
e
n d
at
a
m
at
r
i
x
,
c
l
us
t
er
i
n
g t
ec
hno
l
o
g
y
[
1]
di
v
i
d
es
d
at
a
i
n
t
o s
e
v
er
al
gr
oups
ac
c
or
di
ng
t
o
t
h
e
s
i
m
i
l
ar
i
t
y
of
dat
a.
I
n
t
he
d
i
v
i
s
i
o
n
gr
o
up,
t
h
e
s
i
m
i
l
ar
i
t
y
of
t
he
eac
h
gr
oup
i
n
d
at
a
s
et
i
s
v
er
y
h
i
gh,
at
t
he
s
am
e t
i
m
e,
s
i
m
i
l
ar
i
t
y
of
t
he
dat
a
i
n d
i
f
f
er
ent
gr
ou
ps
i
s
as
s
m
al
l
as
pos
s
i
bl
e.
C
l
us
t
er
i
ng
t
ec
hn
i
q
ue i
s
t
he m
os
t
bas
i
c
r
es
ear
c
h c
ont
en
t
i
n da
t
a m
i
ni
n
g f
i
el
d.
T
her
e ar
e s
om
e c
l
as
s
i
c
al
c
l
us
t
er
i
n
g a
l
gor
i
t
hm
i
nc
l
udi
ng k
-
m
eans
[
2,
3]
,
s
p
ec
t
r
al
c
lu
s
t
e
r
in
g
[
4]
a
nd pr
oba
bi
l
i
t
y
m
odel
i
ng m
et
hods
[
5]
b
as
ed on m
i
x
ed m
odel
.
T
hes
e al
g
or
i
t
hm
s
t
hat
d
i
v
i
d
e dat
a i
n
t
o di
f
f
er
ent
gr
ou
ps
ac
c
or
di
n
g t
o r
o
w
and c
o
l
um
n of
dat
a
m
at
r
i
x
ar
e s
i
ng
l
e
c
l
us
t
er
i
n
g al
gor
i
t
hm
s
.
H
ow
ev
er
,
w
h
en t
h
e di
m
ens
i
on
of
t
he
m
at
r
i
x
i
s
v
er
y
hi
gh
,
c
l
as
s
i
f
i
c
at
i
on
per
f
or
m
anc
e
of
t
he
s
i
ngl
e
c
l
us
t
er
i
n
g
al
gor
i
t
hm
i
s
gr
eat
l
y
r
es
t
r
i
c
t
ed.
W
hen
adopt
i
ng
bi
c
l
us
t
er
i
ng
al
g
or
i
t
hm
t
o
c
l
as
s
i
f
y
d
at
a
i
n
t
he
m
at
r
i
x
,
t
he
s
i
m
i
l
ar
i
t
y
bet
w
e
en
r
o
w
a
nd
c
o
l
um
n
i
n
m
at
r
i
x
s
houl
d
be t
ak
en i
nt
o c
ons
i
der
a
t
i
o
n at
t
he s
am
e t
i
m
e.
T
her
ef
or
e,
i
t
c
an i
m
pr
ov
e
t
h
e per
f
or
m
anc
e of
c
l
us
t
er
i
n
g al
gor
i
t
hm
c
o
m
m
endabl
y
.
I
n r
ec
om
m
end
ed s
y
s
t
em
s
,
e
ac
h r
ow
of
t
he
m
at
r
i
x
r
epr
es
ent
s
a us
er
,
and e
ac
h c
ol
um
n r
epr
es
ent
s
a c
o
m
m
odi
t
y
.
T
r
adi
t
i
on
al
r
ec
om
m
endat
i
on
m
et
hods
t
end t
o f
i
nd
s
i
m
i
l
ar
c
and
i
dat
es
ac
c
or
di
ng t
o
t
he s
i
m
i
l
ar
i
t
y
of
us
er
s
.
I
t
,
ne
v
er
t
h
el
es
s
,
pr
edi
c
t
s
t
hos
e
g
oo
ds
bas
e
d
on
s
i
m
i
l
ar
c
and
i
da
t
es
.
W
i
t
h
b
i
c
l
us
t
er
i
ng
m
et
hod,
i
t
c
an
f
i
nds
s
i
m
i
l
ar
goods
a
nd us
er
s
.
I
n g
ene
ex
pr
es
s
i
on an
al
y
s
i
s
of
bi
oi
nf
or
m
at
i
c
s
,
eac
h r
ow
i
n
a m
at
r
i
x
dat
a
r
epr
es
ent
s
a ge
ne an
d ea
c
h c
ol
um
n r
epr
es
ent
s
a c
ond
i
t
i
on,
s
uc
h as
nor
m
al
s
t
at
e,
a
bnor
m
al
s
t
at
e,
c
anc
er
di
s
e
as
e
s
t
at
e
,
et
c
,
.
Mea
n
w
h
i
l
e
,
bi
c
l
us
t
er
i
ng
al
gor
i
t
hm
c
an
di
v
i
d
e
s
i
m
i
l
ar
gene
i
nt
o
di
f
f
er
ent
gr
oups
on t
he s
i
m
i
l
ar
c
ond
i
t
i
ons
,
w
hi
c
h
c
an
b
et
t
er
a
nal
y
z
e t
he
pat
i
en
t
'
s
p
at
ho
l
og
y
.
T
he
m
ai
n i
dea
of
t
r
adi
t
i
o
na
l
bi
c
l
us
t
er
i
n
g m
et
hod
i
s
t
ha
t
i
t
c
l
us
t
er
s
t
he r
o
w
a
nd c
o
l
u
m
n
o
f
m
at
r
i
x
t
hr
ough t
r
adi
t
i
o
na
l
c
l
us
t
er
i
n
g r
es
pec
t
i
v
e
l
y
bas
ed on t
he s
i
n
gl
e c
l
us
t
er
i
n
g m
et
hod,
and
t
hen
m
er
ges
c
l
us
t
er
i
ng
r
es
ul
t
s
.
T
y
p
i
c
a
l
c
l
us
t
er
i
ng
m
et
hod c
ont
ai
ns
c
ou
pl
e
d
t
w
o
-
w
a
y
c
lu
s
t
e
r
in
g
[6
], f
u
z
z
y
c
-
m
eans
c
l
us
t
er
i
n
g
[
7]
an
d B
i
-
C
or
r
el
a
t
i
o
n C
l
u
s
t
er
i
ng a
l
gor
i
t
h
m
(
B
C
C
A
)
[
8]
et
c
.
I
n or
der
t
o av
oi
d l
i
m
i
t
at
i
ons
of
t
he t
r
adi
t
i
o
nal
c
l
us
t
er
i
ng a
nd b
et
t
er
i
m
pr
ov
e t
he
ef
f
i
c
i
enc
y
of
t
he c
l
us
t
er
i
ng
al
g
or
i
t
hm
,
f
or
ex
a
m
pl
e,
S
B
H
uan
g [
9
]
pr
o
pos
ed
a f
u
z
z
y
c
o
-
c
l
us
t
er
i
ng
a
l
gor
i
t
hm
w
hi
c
h m
i
ni
m
i
z
ed
di
s
t
anc
es
b
et
w
e
en
obj
ec
t
s
and
c
e
nt
er
s
of
c
l
us
t
er
s
i
n
eac
h
f
eat
ur
e
s
pac
e.
Z
F
Y
a
ng
[
1
0]
r
epr
es
ent
e
d
f
u
z
z
y
C
-
m
eans
c
l
us
t
er
i
ng
al
g
or
i
t
hm
b
as
ed
on
t
h
e
i
m
pr
ov
ed
q
uant
um
par
t
i
c
l
e
s
w
ar
m
opt
i
m
i
z
at
i
on.
T
he l
o
c
al
s
ear
c
h ab
i
l
i
t
y
a
nd q
uant
um
gat
es
updat
e s
t
r
at
eg
y
w
er
e i
m
pr
ov
ed
b
y
m
ak
i
ng f
ul
l
us
e
of
t
he ad
v
a
nt
ag
es
of
f
as
t
c
onv
er
g
enc
e
of
quan
t
um
par
t
i
c
l
e s
w
ar
m
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
N
ew
S
emi
-
s
u
per
v
i
s
ed
C
l
us
t
er
i
ng
A
l
gor
i
t
h
m B
as
ed
o
n V
ar
i
at
i
o
n
al
B
ay
es
i
an
…
(
S
h
o
u
lin
Y
in
)
1151
opt
i
m
i
z
at
i
o
n (
Q
P
S
O
)
.
A
nd
Z
hou
Y
[
1
1]
s
ho
w
e
d a s
e
m
i
-
s
uper
v
i
s
e
d m
et
hod t
o i
m
pr
ov
e t
he
c
l
us
t
er
i
n
g r
es
ul
t
s
,
w
hi
c
h
ad
j
us
t
ed t
he s
i
m
i
l
ar
i
t
y
m
at
r
i
x
bas
ed o
n B
a
y
es
i
a
n i
nf
or
m
at
i
on
,
an
d f
i
x
ed
t
he c
l
as
s
l
ab
e
l
s
on t
he r
ef
er
enc
e of
t
h
e p
ai
r
w
i
s
e c
ons
t
r
ai
nt
s
at
l
as
t
.
A
l
t
ho
ugh,
t
hes
e m
et
hods
ar
e
pr
opos
e
d,
t
h
er
e ar
e
s
t
i
l
l
s
o
m
e c
onv
er
genc
e
pr
ob
l
em
s
.
S
o
i
n
t
hi
s
pap
er
,
i
t
pr
es
e
nt
s
a
s
em
i
-
s
uper
v
i
s
ed
c
l
us
t
e
r
i
ng
al
g
or
i
t
hm
bas
ed on
v
a
r
i
abl
e
dec
i
b
el
s
B
a
y
es
i
an.
T
hi
s
ne
w
a
l
gor
i
t
hm
i
s
on t
he bas
i
s
of
t
he
m
at
r
i
x
dat
a,
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
ar
e i
nt
r
o
duc
ed t
o i
t
s
r
ow
s
and c
ol
um
ns
r
es
pec
t
i
v
e
l
y
.
T
he s
t
r
uc
t
ur
e o
f
t
he new
s
c
he
m
e i
s
as
F
i
gur
e
1.
T
he i
nput
d
at
a
i
n t
he
a
l
gor
i
t
hm
i
nc
l
u
des
m
at
r
i
x
dat
a a
nd
aux
i
l
i
ar
y
i
nf
or
m
at
i
on of
c
ol
um
ns
and r
o
w
s
.
A
ux
i
l
i
ar
y
i
nf
or
m
at
i
on
of
c
ol
um
ns
and r
o
w
s
c
an
be r
e
pr
es
ent
e
d b
y
adj
ac
e
nc
y
m
at
r
i
x
of
net
w
or
k
.
T
he c
on
s
t
r
uc
t
i
on
of
t
hi
s
pap
er
i
s
a
s
f
ol
l
o
w
s
.
I
n s
ec
t
i
on
2,
w
e
i
nt
r
oduc
e
t
h
e
m
et
hod of
new
s
c
hem
e,
i
t
c
ont
ai
ns
pr
oba
bi
l
i
s
t
i
c
m
ode
l
and t
he pr
oc
es
s
of
ne
w s
e
m
i
-
s
uper
v
i
s
ed
c
l
us
t
er
i
n
g a
l
gor
i
t
hm
bas
ed
on v
ar
i
a
bl
e dec
i
be
l
s
B
a
y
es
i
an.
I
n
s
ec
t
i
o
n 3,
w
e m
a
k
e ex
per
i
m
ent
s
t
o
s
ho
w
t
h
e ef
f
i
c
i
enc
y
of
our
n
e
w
s
c
hem
e.
S
ec
t
i
o
n 4
gi
v
e
s
a c
onc
l
us
i
on.
F
i
gur
e
1.
S
t
r
uc
t
ur
e of
ne
w
s
c
he
m
e
2.
T
h
e
M
e
th
o
d
o
f
N
ew
S
em
i
-
su
p
er
v
i
sed
C
l
u
s
te
r
i
n
g
A
l
g
o
r
i
th
m
A
s
s
um
i
ng
N
M
R
X
×
∈
is
a
M
×N
-
d
im
e
n
s
io
n
m
a
t
r
ix
.
A
ux
i
l
i
ar
y
i
nf
or
m
at
i
on
of
c
ol
um
n
and
r
o
w i
s
W
h
and
W
c
r
es
pec
t
i
v
e
l
y
.
K
an
d
L
ar
e t
h
e nu
m
ber
o
f
gr
oups
i
n t
he r
o
w
s
and c
ol
um
ns
r
es
pec
t
i
v
el
y
.
K
a
nd
L
ar
e
k
now
n
w
hen m
a
k
i
ng do
ubl
e c
l
us
t
er
i
n
g ana
l
y
s
i
s
.
I
n addi
t
i
o
n,
h
m
r
epr
es
ent
s
t
h
e l
a
t
ent
v
ar
i
ab
l
e of
m
-
th
r
o
w.
z
n
r
e
pr
es
ent
s
t
he l
a
t
ent
v
ar
i
ab
l
e
of
n
-
th
c
ol
um
ns
.
A
nd
)
(
1
M
h
h
h
…
=
,
)
(
1
N
z
z
z
…
=
.
W
e
i
nt
r
oduc
e
G
aus
s
i
a
n d
i
s
t
r
i
but
i
on
2
2
2
)
(
2
/
1
2
2
)
2
(
)
,
|
(
σ
µ
πσ
σ
µ
−
−
−
=
x
e
x
N
,
t
di
s
t
r
i
but
i
on
2
/
1
2
/
2
2
1
}
)
(
1
{
)
(
)
2
/
(
)
2
/
1
2
/
(
)
,
,
(
−
−
−
+
Γ
+
Γ
=
Γ
v
v
x
v
v
v
v
x
µ
λ
π
λ
µ
and
Γ
di
s
t
r
i
but
i
o
n
x
e
x
x
β
α
α
α
β
β
α
−
−
Γ
=
Γ
1
)
(
)
,
,
(
t
o
our
ne
w
al
gor
i
t
hm
.
W
he
r
e
Γ
f
unc
t
i
on i
s
∫
+∞
−
−
=
Γ
0
1
)
(
dt
e
t
x
t
x
i
n t
he
t
di
s
t
r
i
b
ut
i
o
n.
Let
dx
x
d
x
)
(
l
og
)
(
Γ
=
ϕ
,
≠
=
=
y
x
y
x
y
x
0
1
)
,
(
δ
.
I
n t
hi
s
p
ar
t
,
t
h
e r
eas
on
w
h
y
w
e
i
nt
r
od
uc
e G
a
us
s
i
an
di
s
t
r
i
b
ut
i
o
n a
nd
t
di
s
t
r
i
but
i
o
n i
s
t
hat
t
he
y
m
a
k
e
a c
ont
r
i
b
ut
i
on t
o
s
ol
v
e
t
he
f
ol
l
o
w
i
ng
pr
i
or
pr
oba
bi
l
i
t
y
,
p
r
i
or
di
s
t
r
i
but
i
o
n an
d p
os
t
er
i
or
di
s
t
r
i
but
i
on.
2.
1
.
P
r
o
b
a
b
i
l
i
s
ti
c
M
o
d
el
B
ased
o
n
A
u
x
ilia
r
y
In
f
o
r
m
a
ti
o
n
F
i
r
s
t
l
y
,
w
e es
t
ab
l
i
s
h t
he n
e
w
pr
oba
bi
l
i
s
t
i
c
m
odel
f
or
our
s
c
he
m
e.
I
f
m
at
r
i
x
X
,
l
ur
k
i
ng
v
ar
i
ab
l
e
h
an
d
z
ar
e
r
egar
d
ed as
dat
a
i
nf
or
m
at
i
on,
t
he
j
oi
nt
pr
o
bab
i
l
i
t
y
c
an
b
e ex
pr
es
s
ed b
y
t
he
f
ol
l
o
w
i
ng f
or
m
ul
a:
)
|
(
)
|
(
)
,
,
|
(
)
|
,
,
,
(
0
0
θ
θ
θ
θ
θ
p
w
z
p
h
z
X
p
h
z
X
p
h
=
(
1)
W
h
er
e
θ
i
s
p
ar
am
et
er
of
t
he j
oi
nt
di
s
t
r
i
but
i
on
,
0
θ
i
s
s
uper
par
am
et
er
v
ec
t
or
.
U
nder
t
he
c
ondi
t
i
o
n of
k
now
n
θ
,
h
an
d
z
,
c
on
di
t
i
o
nal
di
s
t
r
i
but
i
o
n of
X
c
an be ex
pr
es
s
ed
b
y
G
aus
s
i
an
di
s
t
r
i
b
ut
i
on:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
50
–
1
156
1152
)
)
(
,
|
(
)
|
(
;
)
|
(
)
,
,
|
(
1
1
1
−
=
=
=
=
∏
∏
lk
lk
lk
N
n
z
h
mn
M
m
s
X
N
X
p
X
p
h
z
X
p
n
m
µ
θ
θ
θ
(
2)
W
h
er
e
)
,
(
lk
lk
lk
s
µ
θ
=
.
F
or
m
ul
a(
2)
ai
m
s
t
o e
ns
ur
e t
hat
dou
bl
e c
l
u
s
t
er
el
em
ent
s
c
ont
ai
n
t
h
e
s
a
m
e
v
al
ue
i
n
eac
h
gr
ou
p.
U
nder
t
h
e
c
ond
i
t
i
on
of
k
now
n
z
w
an
d
h
w
,
z
w
and
h
w
i
s
t
he
s
uper
par
am
et
er
s
of
t
he pr
i
or
k
no
w
l
e
dge
.
O
n ac
c
ount
of
t
he f
ac
t
t
hat
aux
i
l
i
ar
y
i
nf
or
m
at
i
o
n i
s
gr
aph
s
t
r
uc
t
ur
e,
w
e
c
an
us
e
M
ar
k
ov
r
andom
f
i
el
d
t
o
c
al
c
ul
at
e
t
h
e
pr
oba
bi
l
i
t
y
of
)
|
(
h
w
h
p
a
n
d
)
|
(
z
w
z
p
r
es
pec
t
i
v
el
y
.
T
her
ef
or
e,
pr
i
or
pr
o
bab
i
l
i
t
y
of
l
ur
k
i
ng v
ar
i
a
bl
e
h
a
nd
z
c
a
n be
r
epr
es
ent
e
d b
y
:
∑
∑
∑
∑
⋅
=
=
=
=
=
+
⋅
M
m
m
L
l
h
l
M
j
j
i
h
ji
h
A
M
i
l
h
w
h
h
W
w
h
h
e
C
w
h
p
1
1
1
1
)
,
(
)
,
(
/
1
)
|
(
δ
δ
(
3)
∑
∑
∑
∑
⋅
=
=
=
=
=
+
⋅
N
n
n
K
k
z
l
N
j
j
i
z
ji
z
A
N
i
k
z
w
z
z
W
w
z
z
e
C
w
z
p
1
1
1
1
)
,
(
)
,
(
/
1
)
|
(
δ
δ
(
4
)
W
h
er
e
h
C
and
z
C
ar
e s
t
and
ar
di
z
ed c
o
ef
f
i
c
i
ent
,
and t
he
y
c
a
n nor
m
al
i
z
e
)
|
(
h
w
h
p
and
)
|
(
z
w
z
p
as
dec
i
m
al
s
bet
w
ee
n
0
an
d
1.
I
f
l
ur
k
i
n
g
v
ar
i
a
bl
e
h
and
z
c
a
n
m
a
k
e
m
or
e
c
onnec
t
i
on
s
har
i
ng t
hr
ou
gh a
ux
i
l
i
ar
y
g
r
aph i
nf
or
m
at
i
on
z
w
and
h
w
,
t
h
en i
t
c
a
n ob
t
ai
n l
ar
ger
c
on
di
t
i
on
al
pr
oba
bi
l
i
t
y
,
w
h
i
c
h
s
ho
w
s
t
h
at
i
t
c
an
i
nt
r
o
duc
e
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
i
nt
o
bi
c
l
us
t
er
i
n
g
al
gor
i
t
hm
b
y
pr
oba
bi
l
i
t
y
d
i
s
t
r
i
b
ut
i
on.
T
hi
s
paper
as
s
u
m
es
t
hat
pr
i
or
di
s
t
r
i
b
ut
i
on
of
par
am
et
er
θ
i
s
c
onj
ugat
e
pr
i
or
di
s
t
r
i
but
i
on,
nam
el
y
:
)
,
|
(
)
)
(
,
(
)
|
(
;
)
|
(
)
|
(
0
0
1
0
0
0
1
0
1
0
β
α
ξ
µ
µ
θ
θ
θ
θ
θ
θ
lk
lk
lk
lk
K
k
lk
L
l
s
G
s
N
p
p
p
−
=
=
=
=
∏
∏
(
5)
2.
2
.
T
h
e
N
ew
S
em
i
-
s
u
p
e
r
v
i
s
e
d
C
l
u
s
te
r
i
n
g
A
l
g
o
r
i
th
m
B
a
s
e
d
o
n
V
ar
i
a
b
l
e D
ec
i
b
el
s
B
a
y
e
si
an
I
n
t
h
i
s
s
ubs
ec
t
i
on
,
w
e
m
ai
n
l
y
i
l
l
us
t
r
at
e
t
he
ne
w
m
et
ho
d
t
hr
o
ugh
t
h
e
i
m
pr
ov
ed
f
or
m
ul
as
.
T
hi
s
paper
ado
pt
s
m
at
r
i
x
X
an
d
0
θ
t
o c
al
c
u
l
at
e
z
,
h
a
nd
θ
i
n f
or
m
ul
a
(
1)
,
t
hen
i
t
us
es
j
oi
nt
di
s
t
r
i
b
ut
i
on t
o c
on
duc
t
d
ou
bl
e c
l
us
t
er
i
ng
ana
l
y
s
i
s
.
H
o
w
ev
er
,
w
h
en us
i
ng
B
a
y
es
i
an m
et
hod t
o
s
t
ud
y
t
he ab
ov
e
par
am
et
er
s
,
t
he c
om
put
at
i
on
al
c
om
pl
ex
i
t
y
i
s
v
er
y
hi
gh.
S
o i
t
ne
eds
a k
i
nd of
appr
ox
i
m
at
e
B
a
y
es
i
a
n
l
ea
r
ni
ng
m
et
hod.
T
hi
s
p
aper
ado
pt
s
V
ar
i
a
t
i
o
na
l
B
a
y
es
m
et
hod
an
d
as
s
u
m
es
t
hat
p
os
t
er
i
or
di
s
t
r
i
but
i
on
of
par
am
et
er
s
ar
e i
ndep
end
ent
of
eac
h
ot
h
er
.
S
o
w
e c
an
get
t
he f
ol
l
o
w
s
:
)
)
(
(
)
)
(
(
)
)
(
(
)
,
,
(
1
1
1
1
∏
∏
∏
∏
=
=
=
=
⋅
⋅
=
M
m
m
N
n
n
K
k
lk
L
l
h
q
z
q
q
h
z
q
θ
θ
(
6)
W
h
er
e
)
(
⋅
q
i
s
pos
t
er
i
or
d
i
s
t
r
i
but
i
on of
V
ar
i
at
i
o
nal
B
a
y
e
s
i
an.
T
hen,
w
e op
t
i
m
i
z
e
V
ar
i
at
i
ona
l
B
a
y
es
i
a
n
di
s
t
r
i
but
i
on
t
hr
o
u
gh m
i
ni
m
u
m
t
r
ue pos
t
er
i
or
di
s
t
r
i
b
ut
i
on
and
K
ul
l
bac
k
-
Lei
b
l
er
di
f
f
er
enc
e
of
V
ar
i
at
i
ona
l
B
a
y
es
i
an
di
s
t
r
i
but
i
on
ac
c
or
di
ng
t
o
t
h
e
k
now
n
i
nf
or
m
at
i
on.
T
h
e
m
in
im
iz
a
t
io
n
m
et
hod i
s
eq
ui
v
a
l
ent
t
o t
he
pr
ev
i
ous
of
m
i
ni
m
i
z
at
i
on
B
a
y
e
s
i
an f
r
ee e
ner
g
y
[1
2
] (
F
[q
])
, n
a
m
e
l
y
:
]
[
)
,
,
(
)
|
,
,
,
(
l
og
)
|
,
,
(
)
|
,
,
,
(
l
og
)
|
(
0
0
0
0
q
F
d
h
z
q
h
z
X
p
h
z
q
d
h
z
X
p
X
F
h
z
h
z
t
=
−
≤
−
=
∫
∑
∑
∫
∑
∑
θ
θ
θ
θ
θ
θ
θ
θ
θ
θ
θ
θ
(
7)
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
N
ew
S
emi
-
s
u
per
v
i
s
ed
C
l
us
t
er
i
ng
A
l
gor
i
t
h
m B
as
ed
o
n V
ar
i
at
i
o
n
al
B
ay
es
i
an
…
(
S
h
o
u
lin
Y
in
)
1153
T
he B
a
y
es
i
an f
r
ee e
ner
g
y
,
w
hi
c
h
i
s
al
s
o c
a
l
l
e
d t
h
e v
ar
i
at
i
o
na
l
s
t
oc
has
t
i
c
c
om
pl
ex
i
t
y
an
d
c
or
r
es
ponds
t
o
a l
o
w
er
bou
nd f
or
t
he
B
a
y
es
i
an
ev
i
d
en
c
e,
i
s
a k
e
y
q
uan
t
i
t
y
f
or
m
o
del
s
e
l
ec
t
i
on.
W
e
as
s
u
m
e t
hat
)
(
k
z
q
z
n
k
n
=
=
,
∑
=
n
k
n
k
z
Z
,
)
(
l
h
q
h
m
l
m
=
=
,
∑
=
m
l
m
l
h
H
.
∑
∑
=
n
mn
k
n
l
m
m
lk
X
z
h
X
,
∑
∑
=
n
lk
k
n
l
m
m
lk
X
z
h
X
2
2
.
S
o
w
e c
ou
l
d
get
t
he p
os
t
er
i
or
d
i
s
t
r
i
bu
t
i
o
n
of
V
ar
i
at
i
o
nal
B
a
y
es
i
an
of
lk
µ
and
lk
s
.
)
,
|
(
)
(
);
,
,
|
(
)
(
lk
lk
lk
lk
lk
lk
lk
lk
lk
lk
lk
s
s
q
T
q
β
α
α
ξ
β
α
µ
µ
µ
Γ
=
=
(
8)
A
nd
lk
α
,
lk
β
,
lk
µ
and
lk
ξ
c
an b
e d
ef
i
ned
as
f
ol
l
o
w
s
:
k
l
lk
Z
H
2
1
0
+
=
α
α
(
9)
}
{
2
1
2
2
2
0
0
0
lk
lk
lk
lk
X
µ
ξ
µ
ξ
β
β
−
+
+
=
(
10)
lk
lk
lk
X
ξ
µ
ξ
µ
+
=
0
0
(
11)
k
l
lk
Z
H
+
=
0
ξ
ξ
(
12)
I
n ad
di
t
i
o
n,
t
h
e p
os
t
er
i
or
di
s
t
r
i
but
i
on
of
V
ar
i
at
i
ona
l
B
a
y
es
i
an
)
(
n
z
q
of
n
z
i
s a
s:
∑
=
=
=
K
k
n
nk
nk
e
e
k
z
q
1
/
)
(
γ
γ
(
13)
n
k
γ
c
an b
e c
al
c
u
l
at
ed
b
y
t
he f
ol
l
o
w
f
or
m
ul
a:
∑
∑
∑
∑
∑
=
=
=
=
=
−
+
−
−
−
−
+
=
L
l
lk
lk
l
L
l
M
m
mn
lk
l
m
lk
lk
L
l
lk
lk
lk
l
N
j
z
k
k
j
z
n
j
z
Z
n
k
H
X
h
H
w
z
W
w
1
1
1
2
1
1
}
l
og
)
(
{
2
1
)
(
2
1
)
1
(
2
1
β
α
ϕ
µ
β
α
α
ξ
α
γ
(
14)
J
us
t
as
t
he
abo
v
e,
pos
t
er
i
o
r
di
s
t
r
i
b
ut
i
on of
m
h
is
:
∑
=
=
=
L
l
m
ml
ml
e
e
l
h
q
1
/
)
(
η
η
(
15)
ml
η
i
s a
s f
o
l
l
o
w
s:
∑
∑
∑
∑
∑
=
=
=
=
=
−
+
−
−
−
−
+
=
K
k
lk
lk
k
K
k
N
n
mn
lk
k
n
lk
lk
K
k
lk
lk
lk
k
M
j
h
l
l
j
h
mj
h
A
ml
Z
X
z
Z
w
h
W
w
1
1
1
2
1
1
}
l
og
)
(
{
2
1
)
(
2
1
)
1
(
2
1
β
α
ϕ
µ
β
α
α
ξ
α
η
(
16)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
50
–
1
156
1154
F
i
na
l
l
y
,
w
e us
e
V
ar
i
at
i
on
al
B
a
y
es
i
a
n m
et
hod t
o ge
t
)
(
θ
q
,
)
(
z
q
a
nd
)
(
h
q
,
and
put
z
,
h
and
θ
i
nt
o f
or
m
ul
a (
1)
t
o an
al
y
z
e
B
i
c
l
us
t
er
i
n
g.
3.
E
xp
e
r
i
m
en
t
al
re
s
u
l
t
s an
d
an
al
y
s
i
s
.
E
x
per
i
men
t
1.
I
n
t
hi
s
pa
per
,
V
ar
i
at
i
on
al
B
a
y
es
i
an
s
em
i
-
s
uper
v
i
s
e
d
B
i
c
l
us
t
er
i
ng
m
et
hod
i
s
abbr
e
v
i
at
ed
as
V
B
S
B
.
W
e
m
a
k
e
c
o
m
p
ar
i
s
on
t
o
B
a
y
es
i
a
n
m
et
hod
(
B
M)
and
k
-
m
eans
m
et
hod.
P
er
f
or
m
anc
e
ev
a
l
u
at
i
on
c
r
i
t
er
i
on
us
es
nor
m
al
i
z
ed
m
ut
ual
i
nf
or
m
at
i
on(
N
MI
)
,
a
l
s
o
w
e
us
e
s
ev
e
r
al
ex
per
i
m
ent
s
t
o i
l
l
us
t
r
at
e t
he n
e
w
al
g
or
i
t
h
m
.
P
ar
am
et
er
s
e
tti
n
g
of
ex
per
i
m
ent
s
i
s
1
=
=
h
l
z
k
w
w
,
2
0
=
α
,
1
0
=
β
,
1
0
=
µ
,
h
A
z
A
w
w
=
.
T
he
y
ar
e c
l
os
el
y
t
o
t
he
t
r
u
e v
al
u
es
r
ang
i
n
g f
r
o
m
an ac
c
ept
ab
l
e
v
a
l
u
e.
W
e c
o
mpar
e t
h
e t
hr
ee
al
g
or
i
t
hm
’
s
per
f
or
m
anc
e of
s
y
nt
h
et
i
c
d
at
a s
e
t
s
.
Let
K
=3
,
L
=2
,
e
ac
h
gr
ou
p
of
bi
c
l
us
t
er
i
n
g
c
ont
a
i
ns
20
r
ow
s
an
d
50
c
ol
um
ns
.
F
or
c
ol
um
n
k
,
)
50
1
)
1
(
50
(
k
n
k
k
z
n
≤
≤
+
−
=
∗
.
F
o
r
ro
w
l
,
)
20
1
)
1
(
20
(
k
m
l
k
h
m
≤
≤
+
−
=
∗
.
A
f
te
r
gener
at
i
ng t
he gr
oups
,
i
t
c
an pr
od
uc
e
m
×n
m
a
t
r
ix
mn
X
b
y
G
aus
s
i
an di
s
t
r
i
b
ut
i
on
)
,
(
2
σ
µ
lk
N
.
W
h
er
e
∗
=
m
h
l
,
∗
=
n
z
k
,
1
1
1
=
µ
,
0
1
2
=
µ
,
1
1
3
=
µ
,
0
2
1
=
µ
,
0
2
2
=
µ
,
5
.
0
2
3
=
µ
.
W
e
us
e
t
he
t
hr
ee
p
ar
am
et
er
s
in
R
,
o
ut
R
and
s
R
t
o
gen
er
at
e
t
he
c
or
r
es
pond
i
n
g
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
net
w
or
k
.
W
her
e
in
R
r
epr
es
ent
s
t
he
pr
opor
t
i
on
of
edge
i
n
one
i
nt
er
n
al
gr
o
up.
o
ut
R
denot
es
t
he
pr
opor
t
i
on
of
edg
e
w
i
t
hi
n
m
i
x
ed
gr
oups
.
s
R
i
s
t
he
pr
op
or
t
i
on
of
nodes
w
i
t
h
l
a
be
l
s
.
F
or
c
ol
um
n
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
z
W
,
i
t
pr
oduc
es
3×
50
×
4
9×
in
R
i
nt
er
n
al
s
i
des
and
2×
50
×
5
0×
o
ut
R
m
i
xe
d
gr
oups
s
i
des
.
F
or
r
o
w
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
h
W
,
i
t
pr
od
uc
es
2×
20
×
1
9×
in
R
i
nt
er
na
l
s
i
de
s
and
20×
20
×
o
ut
R
m
i
x
ed gr
oups
s
i
de
s
.
T
o
r
eal
i
z
e a
s
em
i
-
s
uper
v
i
s
e
d
l
e
ar
ni
ng
a
l
gor
i
t
hm
,
i
t
r
andom
l
y
r
em
ov
es
50
an
d
20
s
i
de
s
f
or
c
ol
um
n
and
r
o
w
r
e
s
pec
t
i
v
el
y
w
i
t
h
no
l
ab
el
s
i
n
nod
e
af
t
er
gener
at
i
ng a
ux
i
l
i
ar
y
n
et
w
or
k
z
W
and
h
W
.
D
ur
i
n
g
t
he
ex
per
i
m
ent
s
,
w
e
s
el
ec
t
(
0.
1,
1,
4)
,
(
0.
15,
1,
3
)
and
(
0.
1,
0.
75,
4)
as
t
he
v
al
u
e
of
(
o
ut
R
,
s
R
,
2
σ
)
.
F
or
eac
h v
a
l
ue
,
i
t
r
an
dom
l
y
pr
oduc
es
1
0 dat
a a
nd s
el
ec
t
s
t
h
e av
er
ag
e v
a
l
ue.
F
i
gur
e
2
and F
i
gur
e
3 i
s
N
MI
per
f
or
m
anc
e c
o
m
par
i
s
on of
t
he r
o
w
an
d c
ol
um
n c
l
us
t
er
i
ng
w
i
t
h t
he
v
a
l
ue (
0.
1
,
1,
4)
.
I
n a
dd
i
t
i
on
,
w
i
t
h t
he i
nc
r
eas
i
n
g of
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
w
ei
ght
,
N
MI
of
V
B
S
B
al
g
or
i
t
hm
i
nc
r
eas
es
f
i
r
s
t
an
d t
he
n d
ec
r
eas
es
.
F
i
gur
e
2.
R
o
w
c
l
us
t
er
i
n
g c
o
m
par
i
s
on w
i
t
h t
he
v
a
l
ue (
0
.
1,
1
,
4)
F
i
gur
e
3.
C
ol
um
n c
l
us
t
er
i
ng
c
om
par
i
s
on w
i
t
h
t
he
v
a
l
ue (
0
.
1,
1,
4)
F
ro
m
F
ig
ur
e
2,
w
e
c
an
k
no
w
t
ha
t
B
M
an
d
k
-
m
eans
m
et
hod
r
eac
hes
a
p
l
at
eau
at
0.
3
8
and
0.
3
r
es
pec
t
i
v
el
y
.
W
hen
w
e
i
gh
t
i
s
0.
1
,
t
he
N
MI
of
V
B
S
B
i
s
ap
pr
ox
i
m
at
el
y
0.
8,
w
h
i
c
h
i
s
t
h
e
hi
g
hes
t
.
S
i
m
i
l
ar
l
y
,
t
he N
MI
of
V
B
S
B
i
s
s
uper
i
or
to
B
M
and k
-
m
eans
i
n F
ig
ur
e
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
N
ew
S
emi
-
s
u
per
v
i
s
ed
C
l
us
t
er
i
ng
A
l
gor
i
t
h
m B
as
ed
o
n V
ar
i
at
i
o
n
al
B
ay
es
i
an
…
(
S
h
o
u
lin
Y
in
)
1155
T
hen w
e s
e
l
ec
t
(
0.
1
5,
1,
3)
and (
0.
1,
0.
75,
4)
as
t
h
e v
al
ue of
(
o
ut
R
,
s
R
,
2
σ
)
.
R
epea
t
t
he
abo
v
e
ex
per
i
m
ent
s
.
A
n
d
w
e get
t
he
F
ig
u
r
e
4,
5,
6,
7.
F
i
gur
e
4.
R
o
w
c
l
us
t
er
i
n
g c
o
m
par
i
s
on w
i
t
h t
he
v
a
l
ue (
0
.
15,
1,
3)
F
i
gur
e
5.
C
ol
um
n c
l
us
t
er
i
ng
c
om
par
i
s
on w
i
t
h
t
he
v
a
l
ue (
0
.
15
,
1,
3)
F
i
gur
e
6.
R
o
w
c
l
us
t
er
i
n
g c
o
m
par
i
s
on w
i
t
h t
he
v
a
l
ue (
0
.
1,
0
.
75
,
4)
F
i
gur
e
7.
C
ol
um
n c
l
us
t
er
i
ng
c
om
par
i
s
on w
i
t
h
t
he
v
a
l
ue (
0
.
1,
0.
75
,
4)
F
ro
m
F
ig
ur
e
4
-
7
w
e
c
an
k
no
w
t
h
at
t
he
y
h
a
v
e
t
he
s
i
m
i
l
ar
r
es
ul
t
s
.
T
he
N
MI
per
f
or
m
anc
e
of
V
B
S
B
a
l
g
or
i
t
hm
obv
i
o
u
s
l
y
ex
c
ee
ds
t
ha
t
i
n
ot
h
er
t
w
o a
l
gor
i
t
hm
s
w
i
t
h r
eac
h
i
ng p
l
at
eau
at
c
ons
t
ant
l
ev
el
.
I
n
F
ig
u
r
e
4,
5,
i
f
t
he
w
ei
g
ht
of
aux
i
l
i
ar
y
i
nf
or
m
at
i
on i
s
hi
gher
,
s
o t
he N
MI
p
er
f
or
m
anc
e
of
V
B
S
B
a
l
g
o
r
i
t
hm
i
s
pr
ed
i
c
t
ed
t
o
ex
per
i
enc
e
a
d
ec
r
eas
i
n
g
t
r
e
nd.
T
her
ef
or
e,
t
he
v
a
l
ue of
2
σ
pl
a
y
s
a s
i
gni
f
i
c
ant
i
nf
l
ue
nc
e on
t
he
al
gor
i
t
hm
.
F
i
gur
e
6
al
s
o s
h
o
w
s
t
hat
N
M
I
per
f
or
m
anc
e c
o
m
par
i
s
on
of
t
he
r
ow
c
l
us
t
er
i
ng
w
i
t
h t
h
e
v
al
ue (
0.
1
5,
1,
3)
i
n
V
B
S
B
i
s
s
uper
i
or
to
B
M
w
h
i
c
h onl
y
ac
c
ount
s
f
or
appr
ox
i
m
at
el
y
0.
5 and
k
-
m
eans
w
i
t
h n
ear
l
y
0
.
4.
T
he anal
y
s
i
s
i
s
s
im
ila
r
t
o
F
i
gur
e
7.
I
n
ad
di
t
i
on,
V
B
S
B
al
gor
i
t
hm
has
a f
as
t
c
onv
er
ge
nc
e r
at
e
.
S
o
t
h
e n
e
w
al
g
or
i
t
hm
has
bee
n pr
o
v
e
d.
E
x
per
i
men
t
2
.
I
n or
d
er
t
o
v
er
i
f
y
t
he
ef
f
i
c
i
enc
y
of
V
B
S
B
a
l
g
or
i
t
hm
,
w
e m
a
k
e an i
nt
r
us
i
o
n d
et
ec
t
i
on
ex
per
i
m
ent
w
i
t
h o
ur
ne
w
m
et
hod and m
ak
e a c
o
m
par
i
s
on
w
i
t
h
r
ef
er
enc
e
[
1
3]
.
T
he
m
et
hod i
n
[
13]
i
s
t
hat
i
t
us
es
p
ar
t
of
m
ar
k
ed dat
a
f
r
o
m
t
he s
am
pl
e da
t
a s
e
t
a
nd
ge
ner
at
es
t
he
S
ee
d s
et
f
or
i
ni
t
i
a
l
i
z
i
ng
t
he c
l
us
t
er
.
B
y
c
al
c
ul
at
i
ng t
he
E
uc
l
i
d
ean
di
s
t
anc
e be
t
w
ee
n m
ar
k
ed poi
nt
i
n s
am
pl
e
dat
a s
e
t
a
nd t
he
av
er
age
v
al
u
e of
l
a
be
l
ed
dat
a i
n e
ac
h
c
l
us
t
er
a
nd
get
t
i
ng
t
h
e i
ni
t
i
al
c
en
t
er
p
oi
nt
,
i
t
ef
f
ec
t
i
v
el
y
a
v
o
i
ds
t
h
e B
l
i
ndnes
s
an
d r
and
om
nes
s
w
hen c
ho
os
i
n
g i
n
i
t
i
a
l
c
l
us
t
er
i
ng c
e
nt
er
b
y
t
r
adi
t
i
on
al
c
l
us
t
er
i
ng
a
l
go
r
i
t
hm
.
W
e
i
nt
r
oduc
e
p
er
f
or
m
anc
e
i
ndi
c
a
t
or
s
t
o
c
o
m
par
e t
he
per
f
or
m
anc
e of
al
gor
i
t
hm
s
i
nc
l
ud
i
n
g det
ec
t
i
o
n r
at
e (
D
R
)
and f
al
s
e pos
i
t
i
v
e r
at
e (
F
P
R
)
.
I
n t
h
is
ex
per
i
m
ent
,
w
e s
el
ec
t
r
epr
es
ent
at
i
v
e 50
00 dat
a.
A
n
d ot
her
dat
a ar
e s
el
ec
t
e
d as
i
n [
13]
.
T
hr
ough
t
es
t
i
n
g
t
he
D
R
an
d
F
P
R
of
5000
a
ggr
es
s
i
v
e
da
t
a
w
i
t
h
di
f
f
er
ent
al
gor
i
t
hm
s
,
w
e
c
an
m
eas
ur
e
t
he
det
ec
t
i
on
ef
f
ec
t
of
eac
h al
g
or
i
t
hm
.
W
e adopt
V
B
S
B
a
n
d t
he m
et
ho
d
i
n
[
13
]
an
d ge
t
t
he
det
ec
t
i
on
r
es
ul
t
s
as
F
i
gur
e
8.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
1
4
,
N
o
.
3
,
S
ept
em
ber
201
6
:
11
50
–
1
156
1156
F
i
gur
e
8.
C
om
par
i
s
on
w
i
t
h
V
B
S
B
an
d r
ef
er
enc
e
[
1
3]
I
t
i
s
c
l
ear
l
y
f
r
om
F
i
gur
e
8
t
hat
D
R
of
V
B
S
B
near
l
y
i
s
89%
ov
er
t
h
at
of
r
ef
er
enc
e
[1
3
]
abou
t
8
7%
.
N
ev
er
t
he
l
es
s
,
f
or
F
P
R
,
V
B
S
B
onl
y
ac
c
oun
t
s
f
or
r
oughl
y
30%
and
r
ef
er
enc
e
[1
3
] w
i
th
35%
.
T
her
ef
or
e,
t
he
r
es
ul
t
s
s
how
t
h
at
t
h
e
ne
w
s
em
i
-
s
uper
v
i
s
ed
c
l
us
t
er
i
n
g
al
gor
i
t
hm
bas
ed on
V
ar
i
at
i
ona
l
B
a
y
es
i
a
n i
s
v
er
y
ef
f
ec
t
i
v
e f
or
t
h
e de
t
ec
t
i
on
w
i
t
h
a l
o
w
er
f
al
s
e
pos
i
t
i
v
e r
at
e.
4.
C
o
n
c
l
u
s
i
o
n
T
hi
s
paper
put
s
f
or
w
ar
d
a
s
em
i
-
s
uper
v
i
s
e
d c
l
us
t
er
i
n
g al
g
or
i
t
hm
bas
ed on
V
ar
i
at
i
o
na
l
B
a
ye
s
i
a
n
.
T
he
al
gor
i
t
hm
not
o
nl
y
c
o
nt
a
i
ns
t
he t
ar
get
m
at
r
i
x
,
but
a
l
s
o
i
nt
r
oduc
e
s
t
he
aux
i
l
i
ar
y
i
nf
or
m
at
i
on
of
r
ow
and
c
ol
u
m
n
i
nt
o
i
t
s
pr
oc
es
s
.
T
he
pr
opos
ed
a
l
g
or
i
t
hm
c
o
m
bi
nes
t
ar
get
m
at
r
i
x
w
i
t
h
t
he
a
ux
i
l
i
ar
y
i
nf
or
m
at
i
on
as
a
j
o
i
nt
d
i
s
t
r
i
but
i
o
n
pr
obab
i
l
i
t
y
m
odel
,
t
h
en
i
t
ad
o
pt
s
V
ar
i
at
i
o
nal
B
a
ye
s
i
a
n
l
e
ar
ni
n
g m
et
hod
t
o es
t
i
m
at
e par
am
et
er
s
i
n
t
hi
s
m
odel
.
A
t
t
h
e
end
of
t
hi
s
p
aper
,
w
e
m
a
k
e ex
per
i
m
ent
s
t
hr
ough
s
y
n
t
het
i
c
dat
a s
et
s
and c
om
par
e t
he V
B
S
B
a
l
g
or
i
t
h
m
t
o B
a
y
es
i
an
m
et
hod and
k
-
m
eans
m
et
hod
t
o
v
er
i
f
y
t
h
e g
ood
p
e
r
f
o
r
m
anc
e of
pr
opos
ed
al
gor
i
t
hm
.
I
n t
he
f
ut
ur
e,
w
e ar
e
ex
pec
t
e
d t
o s
t
ud
y
m
or
e adv
a
nc
ed s
em
i
-
s
uper
v
i
s
ed c
l
us
t
er
i
n
g
al
g
or
i
t
hm
s
t
o
i
m
pr
ov
ed o
ur
m
et
hod an
d a
ppl
y
t
hem
i
nt
o pr
ac
t
i
c
al
en
g
i
ne
er
i
n
g ap
pl
i
c
at
i
ons
.
R
ef
er
en
ces
[1
]
P
et
r
i
ni
F
,
F
e
ng
W
C
, H
o
i
s
i
e
A
, e
t a
l
.
T
he Q
uadr
i
c
s
net
w
o
r
k
(
Q
s
N
et
)
:
hi
gh
-
per
f
o
r
m
anc
e
c
l
us
t
er
i
n
g
t
ec
h
nol
o
gy
.
H
ot
I
n
t
er
c
onn
ec
t
s
9,
200
1
,
I
EEE.
200
1:
0
125.
[2
]
J
i
a
R
Y,
G
u
a
n
YY,
Ya
-
Long L
I
.
P
ar
al
l
el
k
-
m
ea
ns
c
l
us
t
er
i
ng
al
gor
i
t
h
m
ba
s
e
d on
M
apR
edu
c
e
m
od
el
.
C
om
put
er
E
ng
i
ne
er
i
n
g &
D
es
i
gn.
20
14.
[3
]
T
i
anhua Li
u
,
S
hou
l
i
n
Y
i
n.
A
n
I
m
pr
ov
ed
K
-
M
eans
C
l
us
t
er
i
n
g A
l
gor
i
t
hm
f
or
K
a
l
m
an
F
i
l
t
er
.
I
C
I
C
E
x
p
r
e
ss
Let
t
er
s
,
P
ar
t
B
:
A
pp
l
i
c
at
i
ons
.
2
015
;
6(
10)
.
[4
]
Long B
,
Z
ha
ng Z
.
S
p
ec
t
r
a
l c
lus
t
e
r
in
g
f
or
m
u
l
t
i
-
t
y
p
e r
el
at
i
o
nal
dat
a:
U
S
8185
481 B
2
.
2
014.
[5
]
Luc
c
hi
n
i
T
,
D
'
E
r
r
i
c
o G
,
C
ont
i
no F
,
et
a
l
.
T
ow
ar
ds
t
he U
s
e of
E
ul
er
i
a
n F
i
e
l
d P
D
F
M
et
hod
s
f
o
r
Co
m
b
u
s
t
i
o
n M
odel
i
n
g i
n
I
C
E
n
gi
ne
s
.
C
om
put
er
S
i
m
ul
at
i
o
n.
2
014
;
7(
1)
.
[6
]
G
et
z
G
,
Lev
i
ne E
,
D
om
any
E
.
C
oupl
e
d t
w
o
-
w
a
y c
l
u
st
er
i
n
g
of
gen
e m
i
c
r
o
ar
r
ay
dat
a
.
P
r
o
c
eedi
n
gs
of
t
he N
at
i
ona
l
A
c
ade
m
y
of
S
c
i
e
nc
e
s
.
20
00
;
97.
[7
]
Lu
C
,
X
i
ao
S
,
G
u
X
.
I
m
pr
ov
i
ng
f
uz
z
y
C
-
m
eans
c
l
u
s
t
er
i
ng
al
gor
i
t
hm
bas
ed
o
n
a
de
n
s
i
ty
-
i
nduc
e
d
di
s
t
anc
e m
eas
ur
e.
J
our
nal
of
E
ngi
ne
er
i
n
g.
2
014
;
1.
[8
]
B
hat
t
a
c
har
y
a A
,
D
e R
K.
Bi
-
C
or
r
el
at
i
on C
l
u
s
t
er
i
ng A
l
gor
i
t
hm
(
B
C
C
A
)
f
or
det
er
m
i
ni
n
g
a s
et
of
c
o
-
r
egul
a
t
ed gen
es
.
B
i
oi
nf
or
m
at
i
c
s
.
20
09;
25.
[9
]
H
uang S
B
,
Y
ang X
X
,
S
hen L
S
,
e
t
a
l
.
F
u
zzy c
o
-
c
l
us
t
er
i
ng al
gor
i
t
h
m
f
or
hi
gh
-
or
d
er
het
er
ogene
ou
s
dat
a
.
J
our
nal
on
C
om
m
uni
c
at
i
ons
.
201
4.
[1
0
]
Y
ang
Z
F
,
S
hi
H
S
,
S
c
hool
S
E
,
e
t
a
l
.
Fu
z
z
y
C
-
m
eans
c
l
us
t
er
i
ng
al
g
or
i
t
h
m
bas
ed
on
i
m
pr
ov
ed
Q
P
S
O
.
M
oder
n E
l
ec
t
r
o
ni
c
s
T
ec
h
ni
q
ue
.
2014
.
[1
1
]
Z
hou Y
,
W
a
ng Y
,
C
hen D
,
et
al
.
S
e
m
i
-
s
uper
v
i
s
ed s
p
ec
t
r
a
l
c
l
u
s
t
er
i
ng
a
l
gor
i
t
h
m
ba
s
ed
on
bay
es
i
an
dec
i
s
i
on
.
J
our
nal
of
C
om
put
at
i
onal
I
nf
or
m
at
i
on
S
y
s
t
em
s
.
20
15
;
1
1
(4
):
13
33
-
1
342.
[1
2
]
W
a
t
an
abe K
,
S
hi
g
a M
,
W
at
an
abe S
.
U
pper
bo
un
d f
or
v
ar
i
a
t
i
ona
l
f
r
e
e ener
gy
of
B
ay
es
i
a
n net
w
or
k
s
.
M
ac
hi
ne Le
ar
ni
n
g.
2
009
;
75(
2)
:
199
-
2
15.
[1
3
]
X
i
a
Z
G
,
W
a
n
L,
C
ai
S
Y
,
e
t
a
l
.
A
s
e
m
i
-
s
uper
v
i
s
ed
c
l
us
t
er
i
n
g
al
gor
i
t
h
m
or
i
ent
ed
t
o
i
nt
r
u
s
i
o
n
det
e
c
t
i
on
.
J
our
n
al
o
f
S
ha
ndon
g U
ni
v
er
s
i
t
y
.
20
12.
Evaluation Warning : The document was created with Spire.PDF for Python.