T
E
L
KO
MNIK
A
, V
ol
.
17
,
No.
6,
Dec
e
mb
er2
01
9
, p
p
.31
91
~
32
02
IS
S
N: 1
69
3
-
6
93
0
,
accr
ed
ited
F
irst
Gr
ad
e b
y K
em
en
r
istekdikti,
Decr
ee
No: 2
1/E/
K
P
T
/20
18
DOI:
10.12928/TE
LK
OM
N
IK
A
.v
1
7
i
6
.
11510
◼
31
91
Rec
ei
v
ed
O
c
tob
er
20
,
20
1
8
; Rev
i
s
ed
F
eb
r
u
ary
12
,
20
1
9
; A
c
c
ep
t
ed
Ma
r
c
h 1
2
, 2
01
9
Classific
ati
on
o
f
w
e
b servic
es
usin
g d
a
ta m
i
ni
ng
algo
rit
hm
s an
d
imp
ro
ved le
arni
ng
mo
d
el
M.
S
w
a
mi Da
s
*
1
, A. Go
v
ar
d
h
an
2
, D.
V
ija
ya
La
ks
h
mi
3
1
M
a
l
l
a
Re
d
d
y
E
n
g
i
n
e
e
r
i
n
g
Co
l
l
e
g
e
,
H
y
d
e
ra
b
a
d
,
I
n
d
i
a
1,
2
J
a
wa
h
a
rl
a
l
Ne
h
ru
T
e
c
h
n
o
l
o
g
i
c
a
l
Uni
v
e
r
s
i
t
y
Hy
d
e
ra
b
a
d
,
H
y
d
e
ra
b
a
d
,
I
n
d
i
a
3
M
a
h
a
tm
a
G
a
n
d
h
i
I
n
s
ti
tu
te
o
f
Te
c
h
n
o
l
o
g
y
,
Hy
d
e
r
a
b
a
d
,
In
d
i
a
*C
o
rre
s
p
o
n
d
i
n
g
a
u
to
r,
e
-
m
a
i
l
:
m
s
d
a
s
.
5
2
0
@g
m
a
i
l
.c
o
m
1
,
g
o
v
a
rd
h
a
n
_
c
s
e
@y
a
h
o
o
.
c
o
.
i
n
2
,
v
i
j
a
y
a
d
o
d
d
a
p
a
n
e
n
i
@
y
a
h
o
o
.
c
o
m
3
Ab
strac
t
As
p
e
r
th
e
g
l
o
b
a
l
d
i
g
i
ta
l
re
p
o
r
t,
5
2
.
9
%
o
f
th
e
worl
d
p
o
p
u
l
a
ti
o
n
i
s
u
s
i
n
g
th
e
i
n
t
e
rn
e
t
,
a
n
d
4
2
%
o
f
th
e
wo
rl
d
p
o
p
u
l
a
ti
o
n
i
s
a
c
ti
v
e
l
y
u
s
i
n
g
e
-
c
o
m
m
e
rc
e
,
b
a
n
k
i
n
g
,
a
n
d
o
th
e
r
o
n
l
i
n
e
a
p
p
l
i
c
a
ti
o
n
s
.
We
b
s
e
r
v
i
c
e
s
a
re
s
o
ft
ware
c
o
m
p
o
n
e
n
t
s
a
c
c
e
s
s
e
d
u
s
i
n
g
n
e
tw
o
rk
e
d
c
o
m
m
u
n
i
c
a
ti
o
n
s
a
n
d
p
ro
v
i
d
e
s
e
r
v
i
c
e
s
to
e
n
d
u
s
e
r
s
.
So
ft
ware
d
e
v
e
l
o
p
e
rs
p
ro
v
i
d
e
a
h
i
g
h
q
u
a
l
i
ty
o
f
web
s
e
r
v
i
c
e
.
To
m
e
e
t
th
e
d
e
m
a
n
d
s
o
f
u
s
e
r
re
q
u
i
re
m
e
n
ts
,
i
t
i
s
n
e
c
e
s
s
a
ry
fo
r
a
d
e
v
e
l
o
p
e
r
to
e
n
s
u
r
e
q
u
a
l
i
ty
a
rc
h
i
te
c
tu
r
e
a
n
d
q
u
a
l
i
ty
o
f
s
e
r
v
i
c
e
s
.
To
m
e
e
t
th
e
d
e
m
a
n
d
s
o
f
u
s
e
r
m
e
a
s
u
re
s
e
rv
i
c
e
q
u
a
l
i
ty
b
y
th
e
ra
n
k
i
n
g
o
f
w
e
b
s
e
rv
i
c
e
s
,
i
n
th
i
s
p
a
p
e
r,
we
a
n
a
l
y
z
e
d
Q
WS
d
a
ta
s
e
t
a
n
d
fo
u
n
d
i
m
p
o
rta
n
t
p
a
ra
m
e
t
e
rs
a
re
b
e
s
t
p
ra
c
ti
c
e
s
,
s
u
c
c
e
s
s
a
b
i
l
i
ty
,
a
v
a
i
l
a
b
i
l
i
t
y
,
re
s
p
o
n
s
e
ti
m
e
,
re
l
i
a
b
i
l
i
t
y
a
n
d
th
r
o
u
g
h
p
u
t,
a
n
d
c
o
m
p
l
i
a
n
c
e
.
We
h
a
v
e
u
s
e
d
v
a
ri
o
u
s
d
a
ta
m
i
n
i
n
g
t
e
c
h
n
i
q
u
e
s
a
n
d
c
o
n
d
u
c
t
e
d
e
x
p
e
r
i
m
e
n
ts
t
o
c
l
a
s
s
i
f
y
Q
WS
d
a
ta
s
e
t
i
n
t
o
fo
u
r
c
a
te
g
o
ri
c
a
l
v
a
l
u
e
s
a
s
c
l
a
s
s
1
,
2
,
3
,
a
n
d
4
.
Th
e
re
s
u
l
ts
a
r
e
c
o
m
p
a
re
d
wit
h
v
a
r
i
o
u
s
t
e
c
h
n
i
q
u
e
s
ra
n
d
o
m
fo
re
s
t,
a
rti
fi
c
i
a
l
n
e
u
ra
l
n
e
two
rk
,
J
4
8
d
e
c
i
s
i
o
n
tre
e
,
e
x
tr
e
m
e
g
ra
d
i
e
n
t
b
o
o
s
ti
n
g
,
K
-
n
e
a
r
e
s
t
n
e
i
g
h
b
o
r,
a
n
d
s
u
p
p
o
r
t
v
e
c
t
o
r
m
a
c
h
i
n
e
.
M
u
l
ti
p
l
e
c
l
a
s
s
i
fi
e
r
s
a
n
a
l
y
z
e
d
,
a
n
d
i
t
was
o
b
s
e
r
v
e
d
th
a
t
th
e
c
l
a
s
s
i
fi
e
r
te
c
h
n
i
q
u
e
e
Xtr
e
m
e
g
ra
d
i
e
n
t
b
o
o
s
ti
n
g
g
o
t
t
h
e
m
a
x
i
m
u
m
a
c
c
u
r
a
c
y
o
f
9
8
.4
4
%
,
a
n
d
r
a
n
d
o
m
f
o
r
e
s
t
g
o
t
th
e
a
c
c
u
ra
c
y
o
f
9
8
.
1
3
%
.
I
n
fu
tu
re
,
we
c
a
n
e
x
te
n
d
t
h
e
q
u
a
l
i
ty
o
f
web
s
e
r
v
i
c
e
fo
r m
i
x
e
d
a
tt
r
i
b
u
t
e
s
.
Key
w
ords
:
c
l
a
s
s
i
f
i
c
a
ti
o
n
,
d
a
ta
m
i
n
i
n
g
,
QoS,
web
s
e
r
v
i
c
e
s
Copy
righ
t
©
2
0
1
9
Uni
v
e
rsi
t
a
s
Ahm
a
d
D
a
hl
a
n.
All
rig
ht
s
r
e
s
e
rve
d
.
1.
Int
r
o
d
u
ctio
n
T
he
W
W
W
w
as
i
nv
e
nte
d
by
T
i
m
B
er
ne
r
s
Le
e
an
d
a
gr
ou
p
of
me
m
be
r
s
i
n
19
94
,
E
urop
ea
n
P
arti
c
l
e
La
bo
r
at
o
r
y
i
n
G
en
ev
a
,
S
wi
tz
er
l
an
d,
the
i
de
a
of
hy
pe
r
tex
t
i
nf
ormati
on
t
o
k
ee
p
pe
r
s
on
a
l
i
nf
ormat
i
on
.
W
3
me
an
s
a
c
l
i
en
t
progr
am
i
n
on
e
s
y
s
tem
wh
i
c
h
s
tarts
to
d
i
s
pl
ay
o
ne
ob
j
ec
t
by
c
l
i
c
k
i
n
g
an
op
t
i
on
an
d
r
etr
i
ev
e
a
no
th
er
ob
j
ec
t
fr
om
a
no
th
er
s
y
s
tem
whi
c
h
i
s
at
r
e
mo
te
s
erv
er us
i
ng
n
etwo
r
k
c
om
m
un
i
c
at
i
o
n [
1]
.
T
he
web
s
erv
i
c
es
are
s
e
l
f
-
c
on
tai
ne
d
,
l
o
os
el
y
c
ou
p
l
e
d
to
de
s
c
r
i
be
m
od
ul
ar
a
pp
l
i
c
ati
on
s
c
an
be
de
s
i
gn
ed
for
i
nte
r
o
p
erabl
e
bu
s
i
ne
s
s
ap
p
l
i
c
at
i
o
n
s
.
T
he
i
nte
r
n
et
i
s
mo
s
t
l
y
us
ed
B
2
C,
B
2
B
,
e
-
c
om
me
r
c
e,
a
nd
oth
ers
,
t
he
growth
o
f
IT
ac
r
os
s
c
o
mp
an
i
es
wor
l
dwi
de
to
p
erf
orm
b
us
i
n
es
s
ac
ti
v
i
t
i
es
.
W
eb
s
erv
i
c
e
us
es
tec
hn
o
l
og
i
es
l
i
k
e
X
M
L,
W
S
D
L,
UDDI
(
s
erv
i
c
e
di
s
c
ov
erie
s
)
,
an
d
S
O
A
P
prot
oc
ol
s
.
W
e
b
s
erv
i
c
e
(
W
S
)
prov
i
de
d
by
i
nte
r
ne
t
,
ba
s
ed
s
y
s
tem
.
T
h
e
we
b
pa
ge
c
a
n
b
e
ac
c
es
s
ed
by
l
oc
ati
ng
t
he
s
erv
i
c
e
r
eg
i
s
tr
y
.
T
he
e
l
e
me
nts
are
us
ed
t
o
bu
i
l
d
the
web
ap
pl
i
c
at
i
on
on
e
of
t
he
i
mp
ort
an
t
w
eb
c
om
po
n
en
ts
i
s
UDDI;
t
h
e
protoc
o
l
s
are
us
ed
for
s
ea
r
c
hi
ng
a
nd
pu
b
l
i
s
h
i
ng
s
erv
i
c
es
.
W
e
b
s
erv
i
c
es
us
e
S
O
A
arc
hi
tec
t
ure,
wh
i
c
h
h
as
three
are
r
ol
es
,
prov
i
d
er,
r
eq
ue
s
ter,
an
d
r
e
gi
s
tr
y
.
A
s
erv
i
c
e
prov
i
de
r
i
s
th
e
on
e
who
c
an
p
ub
l
i
s
h,
s
ea
r
c
hi
n
g,
a
nd
fi
nd
i
ng
se
r
v
i
c
es
.
T
o
d
i
s
c
ov
er
ap
pl
i
c
ati
on
s
by
UDD
I
i
s
a
r
eg
i
s
tr
y
of
s
erv
i
c
es
,
t
he
r
eq
u
es
tor
i
s
t
he
en
d
us
ers
who
wan
t
to
ac
c
es
s
the
w
eb
s
erv
i
c
es
wh
i
c
h
ar
e
pu
bl
i
s
he
d
by
the
s
erv
i
c
e
prov
i
de
r
.
W
eb
s
erv
i
c
e
(
W
S
)
di
s
c
ov
ery
i
s
t
o
i
de
nt
i
fy
t
he
s
erv
i
c
e
wi
t
h
d
es
c
r
i
pti
o
ns
(
W
S
D
L),
an
d
A
P
Is
of
bu
s
i
ne
s
s
s
erv
i
c
es
. W
S
l
i
nk
s
th
e
us
e
of
ea
c
h
W
S
i
nte
r
f
ac
e.
In
20
1
8,
G
l
o
ba
l
Di
g
i
ta
l
S
ui
t
r
ep
orted,
4
bi
l
l
i
on
pe
o
pl
e
(
ap
prox
i
ma
te
l
y
)
wer
e
us
i
ng
the
i
nte
r
ne
t
ac
r
os
s
the
gl
ob
e
a
mo
n
g
the
t
ota
l
po
pu
l
at
i
on
o
f
7.5
93
b
i
l
l
i
o
n,
i
nte
r
ne
t
us
ers
are
52.9
5%,
i
.e
.,
4.0
2
1
bi
l
l
i
on
,
an
ac
ti
v
e
s
oc
i
a
l
m
ed
i
a
us
e
r
42
.0
9%,
i
.
e.,
3.1
96
bi
l
l
i
o
n.
T
h
e
un
i
qu
e
mo
b
i
l
e
us
ers
are
6
7.6
2
%
,
i
.
e.
,
5
.13
5
b
i
l
l
i
on
ac
ti
v
e
m
ob
i
l
e
us
ers
i
n
th
at
38
.9
5%,
i
.
e.
,
2.9
58
bi
l
l
i
o
n
an
d
a
pp
l
i
c
at
i
o
ns
l
i
k
e
fac
eb
oo
k
,
twi
tte
r
,
a
nd
o
the
r
w
eb
-
ba
s
ed
a
pp
l
i
c
at
i
on
s
.
T
h
e
a
pp
l
i
c
ati
on
s
are
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 1
69
3
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No.
6,
Dec
e
mb
er
20
1
9
:
31
9
1
-
3202
3192
i
nc
r
ea
s
i
ng
v
ery
r
ap
i
dl
y
t
he
de
s
i
gn
er
an
d
de
v
e
l
o
pe
r
to
e
ns
ure
th
e
q
ua
l
i
t
y
de
s
i
gn
an
d
de
v
el
op
m
en
t
of
ap
p
l
i
c
a
ti
o
n
s
whi
c
h
me
e
t
th
e
c
l
i
en
t
d
e
ma
nd
s
of
us
er
s
ati
s
fac
t
i
o
n
of
s
erv
i
c
es
[2]
.
W
eb
s
erv
i
c
e
us
es
a
dy
n
a
mi
c
bu
s
i
ne
s
s
en
v
i
r
o
nm
en
t
an
d
us
er
i
nt
erac
ti
o
ns
,
s
erv
i
c
e
qu
a
l
i
ty
,
an
d
s
ati
s
fac
ti
on
.
F
or
ex
a
mp
l
e,
e
-
c
om
me
r
c
e,
w
eb
s
erv
i
c
e
us
i
n
g
S
O
A
arc
hi
t
ec
ture,
i
nte
r
ac
t
i
on
s
of
ty
pe
an
d
ap
p
l
i
c
t
i
on
s
us
e
d
i
n
l
ar
ge
,
m
ed
i
um
,
an
d
s
ma
l
l
s
erv
i
c
e
us
ers
,
fea
t
ures
wi
th
s
erv
i
c
e
c
om
po
n
en
ts
ha
v
e
pro
pe
r
t
i
es
,
fun
c
ti
o
ns
,
an
d
o
pe
r
at
i
o
ns
[3]
.
Day
t
o
da
y
bu
s
i
n
es
s
ac
ti
v
i
t
i
es
by
web
s
erv
i
c
es
ma
d
e
wi
t
h
q
ua
l
i
ty
of
s
erv
i
c
es
,
web
s
er
v
i
c
e
s
el
ec
ti
on
i
s
th
e
mo
s
t
i
mp
ortant
f
or
the
c
o
ns
um
er
to
ac
c
es
s
ap
pl
i
c
ati
on
s
.
T
he
r
es
t
of
t
h
e
pa
pe
r
de
a
l
s
wi
th
s
ec
t
i
o
n
2
as
r
el
ate
d
wor
k
s
,
s
ec
ti
on
3
is
t
h
e
pro
po
s
ed
a
pp
r
oa
c
h
,
s
ec
ti
o
n
4
prov
i
de
t
he
r
es
ul
ts
an
d
d
i
s
c
us
s
i
on
s
an
d
fi
na
l
l
y
s
ec
t
i
on
5 e
nd
s
w
i
th
a
c
on
c
l
us
i
on
a
nd
f
utu
r
e
s
c
op
e
2.
Rela
t
ed
W
o
r
k
M
A
A
l
mu
l
l
a
e
t
al
.
[4
]
prop
os
ed
a
mo
d
el
to
c
l
as
s
i
fy
i
nto
s
p
ec
i
f
i
c
d
om
a
i
ns
op
era
ted
on
tex
t,
s
i
m
i
l
ar
s
erv
i
c
e
by
t
he
F
uz
z
y
ex
pe
r
t
s
y
s
tem
;
the
r
es
ul
ts
are
c
o
mp
are
d
wi
t
h
o
the
r
me
t
ho
ds
.
T
he
W
S
qu
al
i
ty
da
t
as
et
U
DDI
r
eg
i
s
tr
i
es
ha
v
e
2
05
s
e
r
v
i
c
es
whi
c
h
are
c
l
as
s
i
fi
e
d
i
nto
11
c
l
as
s
es
or
do
m
ai
ns
s
uc
h
as
bu
s
i
n
es
s
,
c
om
mu
ni
c
at
i
on
,
an
d
c
om
mu
n
i
c
at
i
on
an
d
ot
he
r
s
.
Ma
k
hl
ug
h
i
a
n
et
al
.
[5]
us
e
of
CB
A
t
oo
l
th
e
s
erv
i
c
es
o
n
de
ma
n
d,
q
ua
l
i
ty
c
o
ns
tr
ai
nts
,
ex
ec
ut
i
on
ti
me
,
a
nd
ac
c
urac
y
of
s
e
l
ec
ti
on
s
.
Li
mi
tat
i
on
s
do
no
t
k
no
w
t
he
i
m
po
r
ta
nc
e
o
f
s
p
ec
i
fi
c
qu
a
l
i
ty
pa
r
am
ete
r
s
. M
oh
a
nty
et
a
l
.
[6]
us
e
d QW
S
da
t
as
et
c
on
t
ai
ns
w
eb
s
erv
i
c
es
whi
c
h
c
a
n b
e
c
l
as
s
i
f
i
ed
i
nto
fou
r
c
a
teg
oric
al
v
al
ue
s
us
i
ng
Ma
r
k
ov
B
l
an
k
et,
Na
i
v
e
B
ay
es
,
an
d
T
a
bu
s
e
arc
h,
th
e
W
S
RF
fr
om
da
ta
,
n
ai
v
e
B
ay
es
i
s
85
.6
2%
us
i
ng
Q
W
S
da
ta,
an
d
oth
e
r
s
m
eth
o
ds
us
e
d.
T
he
l
i
mi
t
ati
on
s
are
a
q
ua
l
i
ty
mo
d
el
av
ai
l
ab
l
e,
bu
t
pred
i
c
ti
o
n
ac
c
urac
y
i
s
l
ow
.
C
h
en
L
i
.
et
a
l
.
[
7]
u
s
e
of
m
od
e
l
s
Nai
v
e
B
ay
e
s
,
s
up
po
r
t
i
n
g
v
ec
tor
ma
c
hi
ne
3
91
we
b
s
erv
i
c
es
i
nto
s
erv
i
c
e
c
l
as
s
i
f
i
c
ati
on
s
us
i
n
g
r
ou
gh
s
et
the
ory
c
l
as
s
i
f
i
c
a
ti
on
of
w
eb
pa
ge
s
i
nt
o
ni
n
e
d
i
ffe
r
en
t
c
l
as
s
es
l
i
k
e
e
d
uc
ati
o
n,
foo
d,
ec
on
om
y
an
d
wea
po
ns
,
l
i
m
i
tat
i
on
s
do
n
ot
pr
ov
i
de
th
e
be
s
t q
u
al
i
ty
of
web
s
erv
i
c
e
.
T
he
G
u
os
he
n
g
K
a
ng
et
a
l
.
[8]
us
e
of
c
o
l
l
a
bo
r
at
i
v
e
m
od
e
l
f
i
l
teri
n
g
(
CF
)
i
s
a
me
t
ho
d
to
predi
c
t
t
he
i
nte
r
es
t
of
us
ers
,
c
ho
i
c
e,
pref
erenc
e,
l
i
k
es
,
an
d
d
i
s
l
i
k
es
.
In
CF
ap
pro
ac
h
the
r
e
ar
e
three
c
o
nc
ep
ts
f
i
r
s
t,
fu
nc
ti
on
a
l
r
e
l
at
i
on
s
(
k
ey
wor
ds
,
i
np
ut,
an
d
o
utp
ut),
s
ec
on
d
i
s
the
s
c
ore
of
the
c
os
i
n
e
s
i
m
i
l
arit
y
me
tr
i
c
s
of
the
us
ers
,
an
d
th
i
r
d
i
s
t
he
ut
i
l
i
ty
op
era
ti
on
s
the
Q
o
S
i
nt
o
h
i
gh
an
d
l
ow
v
al
ue
s
.
M
oh
a
n
P
a
tr
o,
e
t
al
.
[9
]
us
ed
c
l
as
s
i
f
i
ers
to
c
l
as
s
i
fy
the
W
S
on
Q
W
S
da
t
a
s
et
tha
t
ar
e
F
uz
z
y
r
el
ate
d
t
ec
hn
i
qu
es
wi
th
f
ea
t
ure
s
el
ec
ti
o
n,
G
ai
n
r
ati
o
a
n
d
Inf
ormat
i
on
g
ai
n
w
i
th
three
me
th
od
s
w
hi
c
h
are
c
om
p
ared.
H
us
s
ei
n
A
l
-
He
l
a
l
e
t
a
l
.
[
10
]
propos
e
d
an
al
go
r
i
t
hm
r
ep
arab
i
l
i
ty
as
a
me
tr
i
c
to
de
t
ermi
ne
the
web
s
erv
i
c
es
p
l
a
ns
eq
ua
l
or
mo
r
e
to
l
erant
p
l
an
s
.
T
o
di
s
c
ov
er
a
nd
re
-
us
e
web
s
erv
i
c
es
i
n
t
he
organ
i
z
ati
on
to
s
e
l
ec
t
the
s
erv
i
c
es
whi
c
h
are
bu
s
i
ne
s
s
an
d
qu
al
i
ty
o
f
s
erv
i
c
e (
Q
oS
)
ne
ed
s
. Th
e
Q
W
S
da
tas
e
t
i
s
us
e
d
for
ex
pe
r
i
m
en
ts
.
RK
Mo
ha
n
ty
et
al
.
[
11
]
pr
op
os
ed
W
S
c
l
as
s
i
f
i
c
ati
on
us
i
n
g
P
NN,
B
P
N
N,
T
r
ee
n
et,
G
MDH,
S
V
M,
J
48
,
CA
RT
to
pr
ed
i
c
t
qu
a
l
i
ty
,
i
d
en
t
i
fy
an
d
m
ea
s
ure
the
qu
al
i
ty
an
d
us
er
s
a
ti
s
fac
ti
on
.
D
A
A
de
ni
y
i
et
al
.
[12
]
us
ed
m
eth
o
ds
of
K
NN,
CA
RT
,
n
e
ural
n
etw
ork
(
NN)
mo
d
el
i
n
the
s
tud
y
.
T
he
RS
S
r
ea
de
r
s
’
d
ata
c
l
a
s
s
l
ab
el
s
,
w
eb
s
i
t
e,
d
ata
c
at
eg
oric
a
l
v
al
ue
s
wor
l
d,
b
us
i
ne
s
s
,
po
l
i
t
i
c
s
,
s
po
r
ts
,
etc
.
W
S
c
l
as
s
i
fi
c
a
ti
on
J
.
L
i
u
et
a
l
.
[13
]
are
us
i
ng
n
ai
v
e
B
ay
es
s
e
ma
n
ti
c
web
t
o
d
es
c
r
i
be
an
att
r
i
bu
t
e
of
W
eb
s
erv
i
c
e
us
i
ng
me
t
ho
d
da
ta
prep
arati
o
n,
c
l
as
s
i
fy
i
ng
of
O
W
LS
T
C
d
ata
s
et,
s
em
an
t
i
c
web
i
nto
s
ev
en
di
ffe
r
en
t
areas
.
T
h
e
he
uris
t
i
c
ap
pro
ac
h
p
r
op
os
ed
by
M
Ma
k
h
l
u
gh
i
an
[
14
]
by
pr
e
-
proc
es
s
i
ng
,
c
l
as
s
i
fi
c
a
ti
o
n
ac
c
ordi
ng
to
Q
o
S
l
ev
el
s
of
c
an
d
i
d
ate
s
erv
i
c
e
an
d
r
a
nk
i
ng
an
d
s
e
l
ec
ti
ng
th
e
be
s
t
s
erv
i
c
e.
Cl
as
s
as
s
oc
i
ati
o
n
r
ul
es
us
i
n
g
Q
W
S
d
ata
s
et
wi
th
non
-
f
un
c
ti
on
a
l
a
nd
s
ec
urit
y
pa
r
a
me
t
ers
are n
ot
a
d
dre
s
s
ed
.
Q
ua
l
i
ty
ne
ed
f
or
no
n
-
fu
nc
ti
on
al
we
b
s
erv
i
c
e,
i
n
th
e
r
ea
l
-
wor
l
d
da
t
as
et,
21
35
8
web
s
erv
i
c
es
,
ov
er
30
m
i
l
l
i
on
r
e
al
-
wor
l
d
we
b
s
erv
i
c
es
by
v
a
r
i
ou
s
c
ou
n
ti
es
,
fai
l
ure
du
e
t
o
s
om
e
of
th
e
c
au
s
es
l
i
k
e
HT
T
P
ba
d
r
e
qu
es
t,
s
erv
er
err
or,
ba
d
ga
tew
ay
,
s
erv
i
c
e
no
t
a
v
a
i
l
ab
l
e,
ne
tw
ork
un
-
r
ea
c
ha
bl
e
,
c
on
ne
c
ti
on
r
efu
s
ed
,
ti
me
ou
t
an
d
th
e
Res
po
ns
e
ti
me
of
us
ers
[
1
5].
S
ou
ma
d
i
p
G
ho
s
h
et
a
l
.
[1
6]
m
eth
o
d
ne
uro
f
uz
z
y
c
l
as
s
i
fi
er
i
n
pu
t
v
ec
tor,
fuz
z
i
fi
c
at
i
on
i
nt
o
a
r
ti
fi
c
i
al
n
eu
r
a
l
ne
twork
(
A
NN)
an
d
c
l
as
s
i
fi
er
i
nt
o
de
fuz
z
i
f
i
c
at
i
on
.
E
x
pe
r
i
me
nta
l
r
es
u
l
ts
by
the
us
e
of
UCI
r
e
po
s
i
t
ory
da
t
as
et
K
D
D,
breas
t
c
an
c
er
,
i
r
i
s
,
a
nd
oth
er
d
ata
s
ets
f
i
n
d
the
c
l
as
s
i
fi
c
at
i
on
ac
c
urac
i
e
s
.
W
eb
s
erv
i
c
e
d
es
i
g
n,
c
o
ns
um
ers
,
r
e
-
us
e
fu
nc
ti
o
na
l
i
ty
.
A
l
i
O
un
i
etc
.
[17
]
pr
o
po
s
ed
a
hy
brid
ap
proac
h
wh
i
c
h
us
e
s
he
uris
t
i
c
-
ba
s
ed
ap
proac
h
to
the
de
s
i
gn
q
ua
l
i
ty
of
w
eb
i
nt
erfac
es
.
T
he
ex
pe
r
i
me
n
tal
r
es
u
l
ts
c
on
du
c
t
ed
26
r
ea
l
-
wor
l
ds
A
ma
z
on
a
nd
Y
ah
o
o.
W
e
b
s
erv
i
c
e
c
l
as
s
i
fi
ed
by
s
el
ec
ti
on
,
di
s
c
ov
ery
,
an
d
c
om
p
os
i
ti
on
.
T
he
Y
i
l
on
g
Y
an
g
et
al
.
[18
]
prop
os
ed
a
de
e
p
n
eu
r
a
l
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Cl
as
s
i
f
i
c
ati
on
o
f we
b s
erv
i
c
es
us
i
ng
da
t
a m
i
n
i
ng
al
go
r
i
t
hm
s
... (
M.
S
w
am
i
Das
)
3193
ne
twork
s
erv
i
c
e
c
l
as
s
i
f
i
c
ati
on
a
pp
r
o
ac
h.
T
h
e
10
0
00
r
e
al
-
wor
l
d
s
we
b
s
erv
i
c
es
i
nt
o
50
c
ate
g
orie
s
of
v
al
ue
s
. E
v
e
l
u
m
S
eto
an
i
et
a
l
.
[1
9]
us
e
r
e
-
us
ab
l
e t
ex
t c
l
as
s
i
fi
c
a
ti
o
n i
n
B
ah
as
a
In
do
ne
s
i
a.
W
eb
s
erv
i
c
es
de
ma
nd
,
prov
i
de
go
o
d
s
ol
u
ti
o
ns
wi
th
t
he
i
nte
r
o
pe
r
ab
l
e
property
.
Mu
l
t
i
fac
e
ted
m
atc
h
ma
k
i
n
g
fr
am
ework
,
w
eb
s
erv
i
c
e
qu
al
i
ty
,
an
d
us
er
s
e
ek
s
be
s
t
q
ua
l
i
t
y
s
erv
i
c
es
.
S
a
mb
as
i
v
am
et
a
l
.
[
20
]
i
d
en
t
i
f
i
ed
21
q
ua
l
i
ty
pa
r
am
ete
r
s
,
a
nd
ex
pe
r
i
me
nts
c
on
d
uc
ted
10
00
we
b
s
erv
i
c
es
,
s
erv
i
c
e
d
i
s
c
ov
ery
an
d
s
ea
r
c
h.
X
i
on
g
et
a
l
.
[
21
]
prop
os
ed
a
no
v
e
l
de
e
p
l
ea
r
n
i
n
g
hy
bri
d
a
pp
r
oa
c
h
for
web
ap
pl
i
c
ati
on
r
ec
o
m
me
nd
ati
on
an
d
i
mp
r
ov
i
ng
pe
r
forma
nc
e.
W
eb
-
b
as
ed
ap
p
l
i
c
a
ti
on
s
i
n
c
r
ea
s
e
r
ap
i
dl
y
v
ery
fas
t.
T
he
d
es
i
gn
er
an
d
d
ev
el
o
pe
r
en
s
ure
to
prov
i
de
hi
gh
qu
a
l
i
ty
s
erv
i
c
e
s
fo
r
c
us
tom
er s
at
i
s
fac
ti
on
.
P
r
ob
l
em
de
f
i
n
i
t
i
on
:
T
he
s
oft
war
e
de
v
e
l
o
pe
r
a
i
m
i
s
to
d
es
i
gn
an
d
d
ev
el
op
t
h
e
be
s
t
ap
p
l
i
c
at
i
on
wh
i
c
h
w
i
l
l
me
et
th
e
us
er
s
pe
c
i
f
i
c
at
i
on
s
(
i
nc
l
ud
i
ng
fun
c
t
i
on
al
an
d
no
n
-
f
un
c
ti
on
a
l
pa
r
am
ete
r
s
)
ac
c
ordi
n
g
to
t
he
s
erv
i
c
e
l
ev
el
a
gree
me
nt
(
S
LA
)
.
W
eb
s
erv
i
c
e
qu
a
l
i
t
y
i
s
me
as
ured
by
the
r
an
k
i
n
g
of
the
w
eb
ap
pl
i
c
ati
on
.
T
h
e
r
an
k
i
n
g
of
web
s
erv
i
c
es
us
i
ng
c
l
a
s
s
i
fi
c
ati
on
an
d
predi
c
t
i
on
s
.
W
e
ha
v
e
us
ed
Q
W
S
da
t
as
et
f
or
c
on
d
uc
ti
ng
th
e
ex
p
erim
en
ts
t
o
fi
nd
web
s
erv
i
c
e
qu
a
l
i
ty
us
i
ng
v
ario
us
da
ta
mi
ni
ng
me
th
od
s
(
i
.e.
,
c
l
as
s
i
fi
c
a
ti
o
n
a
nd
pred
i
c
ti
o
ns
)
.
T
he
c
l
as
s
i
f
i
c
at
i
on
of
w
eb
s
erv
i
c
e
he
l
ps
th
e
s
oft
w
are
de
s
i
gn
er
to
i
m
prov
e
qu
a
l
i
ty
an
d
pe
r
forma
nc
e.
T
h
e
pro
po
s
e
d
mo
de
l
i
s
s
ho
w
n
i
n
F
i
g
ure
1.
T
h
i
s
i
s
ap
pl
i
e
d
for
th
i
s
probl
e
m
to
s
ol
v
e a
n
d we
b s
erv
i
c
e c
l
as
s
i
fi
c
at
i
on
s
us
i
ng
i
n
pu
t
da
t
as
et
.
F
i
gu
r
e
1.
Q
u
al
i
ty
of
w
eb
s
er
v
i
c
e
c
l
as
s
i
fi
c
a
ti
o
n
l
ea
r
n
i
ng
m
od
e
l
3.
P
r
o
p
o
se
d
Ap
p
r
o
a
ch
T
he
i
nt
ernet
-
b
as
ed
a
pp
l
i
c
a
ti
on
s
qu
al
i
ty
i
s
c
on
s
i
d
ered
by
non
-
fun
c
t
i
on
al
p
aram
et
ers
for
s
ati
s
fy
i
ng
th
e
us
er
r
eq
u
i
r
e
me
nts
w
i
t
h
fu
nc
ti
o
na
l
p
ara
me
ters
.
T
h
e
propos
e
d
m
o
de
l
i
s
s
ho
wn
i
n
ab
ov
e
F
i
gu
r
e
1
.
Th
e
d
ata
mi
n
i
n
g
al
go
r
i
t
hm
s
are
us
e
d
to
c
l
as
s
i
fy
the
q
ua
l
i
ty
of
w
eb
s
erv
i
c
es
,
i
.
e.
Q
W
S
da
t
as
et
.
Ini
ti
a
l
s
t
ag
e
us
e
pre
-
proc
es
s
i
ng
,
s
el
ec
t
i
on
o
f
tr
a
i
n
i
ng
d
ata
wi
t
h
a
t
es
ti
ng
d
ata
s
et
wi
th
us
e o
f v
ari
ou
s
c
l
as
s
i
f
i
c
ati
o
n
m
eth
od
s
us
e
d t
o
fi
nd
a rank
i
ng
of
w
eb
s
erv
i
c
es
.
Le
t t
he
da
tas
e
t
c
on
s
i
s
t
of
A
1
,
A
2
,
..,
A
n
ea
c
h
whi
c
h
b
el
on
gs
to
c
l
as
s
C
i
,
where
C
i
in
{C
1
,
C
2
,..,
C
n
}
where
C
i
>
=
2.
Q
ua
l
i
ty
i
s
m
os
t
i
m
po
r
ta
nt,
whi
c
h
c
o
mp
are
d
w
i
th
att
r
i
b
ute
s
l
i
k
e
l
oa
d
di
s
tr
i
bu
t
i
o
n,
s
erv
i
c
e
di
r
ec
t
i
on
,
throug
hp
u
t,
c
os
t,
r
es
po
n
s
e
ti
me
,
an
d
oth
er
e
l
e
me
nts
.
F
or
ex
am
p
l
e,
e
-
c
om
me
r
c
e
web
ap
p
l
i
c
at
i
on
s
prov
i
d
e
fu
nc
ti
o
na
l
i
ty
as
p
er
the
S
LA
wi
t
h
s
ati
s
fy
i
ng
qu
al
i
ty
pa
r
am
et
e
r
s
.
T
he
Q
W
S
da
tas
et
[2
2
]
was
r
e
l
ev
an
t
ob
j
ec
ts
i
n
th
e
do
m
ai
n.
I
n
thi
s
c
as
e,
d
ata
c
on
ta
i
ns
v
ario
us
qu
al
i
ty
pa
r
am
ete
r
s
s
uc
h
as
r
es
po
ns
e
ti
me
,
throu
gh
pu
t,
av
ai
l
ab
i
l
i
ty
,
ac
c
es
s
i
b
i
l
i
ty
,
r
el
i
ab
i
l
i
ty
,
b
es
t
prac
ti
c
es
,
c
om
pl
i
a
nc
e,
l
at
en
c
y
,
an
d
d
oc
um
e
nta
t
i
o
n.
T
he
c
l
as
s
i
fi
c
a
ti
o
n
of
w
e
b
s
erv
i
c
es
are
Cl
as
s
-
1
(
hi
gh
qu
a
l
i
ty
)
,
C
l
as
s
-
2
m
(
qu
a
l
i
ty
)
,
C
l
as
s
-
3
(
av
erage
qu
a
l
i
ty
)
,
an
d
Cl
as
s
-
4
(
po
or
q
ua
l
i
ty
s
erv
i
c
es
)
.
T
hi
s
i
s
a
pp
l
i
e
d
f
or
th
i
s
pro
bl
em
to
s
ol
v
e
a
nd
w
eb
s
erv
i
c
e
c
l
as
s
i
fi
c
a
t
i
on
s
us
i
ng
In
pu
t
da
tas
et
,
l
ea
r
ni
n
g
me
t
ho
d
s
(
c
l
as
s
i
fi
c
ati
on
tec
h
ni
q
ue
s
s
uc
h
as
r
an
do
m
fores
t,
arti
fi
c
i
al
ne
ura
l
ne
twork
,
J
48
d
ec
i
s
i
on
tr
ee
,
eX
tr
e
me
gra
di
e
nt
b
oo
s
ti
n
g
,
K
-
ne
ares
t
ne
i
g
hb
or
a
nd
s
up
po
r
t
v
ec
tor
ma
c
hi
ne
)
are
us
ed
,
a
nd
fea
ture
s
el
ec
ti
o
ns
l
i
k
e
r
es
po
ns
e
ti
me
,
ac
c
es
s
i
bi
l
i
ty
,
r
e
l
i
ab
i
l
i
ty
,
throug
hp
u
t,
av
a
i
l
a
bi
l
i
ty
,
c
o
mp
l
i
a
nc
e,
an
d
be
s
t
prac
ti
c
e
s
an
d
tr
ai
ni
ng
d
ata
t
o
c
l
as
s
i
fy
the
da
ta
i
nto
c
ate
go
r
i
c
al
v
al
ue
s
(
Ra
nk
1
,
Ra
nk
2,
Ra
nk
3
an
d R
an
k
4
)
.The q
ua
l
i
ty
of
we
b
s
erv
i
c
e
Q
W
S
da
t
as
et
c
on
s
i
s
ts
of
25
0
7
s
a
mp
l
es
us
i
ng
c
l
as
s
i
f
i
c
ati
on
ap
proa
c
he
s
to
fi
n
d
t
he
ac
c
urac
y
of
c
l
as
s
i
f
i
c
at
i
on
me
th
od
s
.
F
i
n
d
th
e
de
ns
i
ty
of
fea
t
ure
att
r
i
bu
tes
r
es
po
n
s
e
ti
me
i
s
s
k
ewed
l
eft
,
av
ai
l
ab
i
l
i
ty
i
s
r
i
g
ht
s
i
de
s
k
ewe
d,
thro
ug
hp
ut
i
s
i
nc
r
ea
s
i
ng
an
d
grad
ua
l
l
y
de
c
r
ea
s
i
ng
,
s
uc
c
es
s
ab
i
l
i
t
y
r
i
gh
t
s
i
de
s
k
ewed
,
t
he
r
el
i
a
bi
l
i
ty
of
c
urv
es
,
c
om
pl
i
a
nc
e
c
urv
es
,
be
s
t
prac
t
i
c
es
at
th
e
r
a
ng
e
an
d
c
l
as
s
l
ab
el
e
d
r
an
k
ed
as
1,
2,
3
a
nd
4,
are
s
ho
w
n
i
n
F
i
gu
r
e
2
.
T
he
Q
W
S
d
ata
r
ea
d
i
n
pu
t,
i
n
i
n
i
ti
al
s
tag
e
pre
-
proc
es
s
i
ng
,
tr
ai
ni
n
g
(
l
a
be
l
l
ed
da
t
a)
wi
t
h
tes
t
i
ng
(
i
n
pu
t
d
ata
)
us
i
ng
l
ea
r
n
i
n
g
m
od
e
l
s
s
uc
h
as
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 1
69
3
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No.
6,
Dec
e
mb
er
20
1
9
:
31
9
1
-
3202
3194
r
an
do
m
fores
t
(
RF
)
,
A
r
t
i
fi
c
i
al
Ne
ural
Netw
ork
,
J
48
D
ec
i
s
i
on
T
r
e
e,
e
X
tr
e
me
gra
di
e
nt
bo
os
ti
n
g
,
K
NN
c
l
as
s
i
fi
er,
an
d
S
V
M
me
th
od
s
are
us
ed
to
c
l
as
s
i
fy
an
d
pred
i
c
t
tes
t
i
ng
da
t
as
et
i
nto
c
l
as
s
l
ab
el
s
(
Cl
as
s
-
1,
2,
3
a
nd
4)
for
the
ex
pe
r
i
me
nt
al
r
es
ul
ts
c
o
nd
uc
t
ed
us
i
n
g
R
progr
am
m
i
n
g.
T
he
de
tai
l
s
of
the
al
go
r
i
t
hm
s
are di
s
c
us
s
ed
be
l
ow.
3.1.
Rand
o
m F
o
r
es
t
T
he
r
a
nd
o
m
fores
t
i
s
a
de
c
i
s
i
on
tr
e
e
w
hi
c
h
h
as
a
c
ol
l
e
c
ti
on
of
de
c
i
s
i
o
n
tr
e
es
k
no
wn
as
fores
t,
the
ne
w
ob
j
ec
t
att
r
i
bu
te
c
l
as
s
i
fi
c
at
i
o
n
of
t
he
c
l
as
s
l
ab
el
,
1,
2,
3,
a
nd
4
.
T
he
fores
t
i
s
c
ho
s
en
c
l
as
s
i
f
i
c
ati
on
ha
s
the
mo
s
t
v
ot
es
ov
era
l
l
tr
ee
i
n
th
e
fores
t.
A
l
g
orit
hm
1
ex
p
l
a
i
ns
r
an
do
m f
ores
t
.
A
l
g
orit
hm
1.
Ra
nd
o
m For
es
t
Input:
QWS dataset
Output:
Classification label class
-
1,2,3 and 4
Begin
Step 1:
The training set is N and sample training set in a growing tree
Step 2:
If input
variables m< M each node, m is variable, M is a best to
split node.
Step3:
Find the Class label in the best split criteri
on leaf node.
Step 4:
Each tree grew at the maximum possible extent; then there will be no
pruning.
End
F
i
gu
r
e
2.
Q
W
S
da
ta
de
ns
i
t
y
v
al
ue
s
RF
c
l
as
s
i
fi
c
a
ti
o
n,
he
r
e
w
e
h
av
e
us
ed
i
n
ex
pe
r
i
me
nts
5
00
tr
e
es
,
s
pl
i
t
e
ac
h
v
ari
ab
l
e
by
2,
ou
t
of
ba
g
err
or
(
O
O
B
es
ti
ma
te
err
or)
v
al
ue
i
s
1.7
9
%
us
i
ng
R
progr
am
m
i
n
g
.
T
he
r
es
ul
ts
of
r
an
do
m
for
es
t
c
on
f
us
i
on
m
atri
x
are
s
ho
wn
i
n
th
e
T
a
bl
e
s
1
a
nd
2.
I
t
g
i
v
es
th
e
c
o
mp
arat
i
v
e
s
tu
dy
of
pred
i
c
ted
c
l
as
s
v
aria
bl
es
wi
th
ob
s
erv
at
i
on
s
.
T
ab
l
e
1.
R
an
do
m F
ores
t
-
c
on
fus
i
o
n
M
atri
x
P
r
e
d
ic
t
ion
R
a
n
k
1
R
a
n
k
2
R
a
n
k
3
R
a
n
k
4
C
las
s
E
r
r
o
r
C
las
s
-
1
583
2
2
6
0
.
0
1
6
8
6
3
4
0
6
C
las
s
-
2
3
803
0
0
0
.
0
0
3
7
2
2
0
8
4
C
las
s
-
3
1
1
587
10
0
.
0
2
0
0
3
3
3
8
9
C
las
s
-
4
2
5
13
489
0
.
0
3
9
2
9
2
7
3
1
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Cl
as
s
i
f
i
c
ati
on
o
f we
b s
erv
i
c
es
us
i
ng
da
t
a m
i
n
i
ng
al
go
r
i
t
hm
s
... (
M.
S
w
am
i
Das
)
3195
T
ab
l
e
2.
C
om
pa
r
at
i
v
e
Res
u
l
ts
w
i
t
h A
c
t
ua
l
V
al
u
es
as
nd
P
r
ed
i
c
ted
C
l
as
s
es
Us
i
ng
R
an
d
om
F
ores
t A
l
g
o
r
i
thm
Us
i
ng
Ten
-
f
o
l
d
Cr
os
s
-
V
al
i
d
ati
on
P
r
e
d
ic
t
ion
Ob
s
e
r
v
a
t
ion
R
a
n
k
1
R
a
n
k
2
R
a
n
k
3
R
a
n
k
4
r
o
w
I
n
d
e
x
Mt
r
y
R
e
s
a
mple
C
las
s
1
R
a
n
k
1
0
.
9
0
0
0
.
0
0
2
0
.
0
3
2
0
.
0
6
6
15
1
Fold
0
1
C
las
s
4
R
a
n
k
4
0
.
0
6
2
0
.
1
3
0
0
.
0
8
6
0
.
7
2
2
16
1
Fold
0
1
C
las
s
3
R
a
n
k
3
0
.
0
3
2
0
.
0
0
0
0
.
8
9
0
0
.
0
7
8
17
1
Fold
0
1
C
las
s
4
R
a
n
k
4
0
.
0
0
4
0
.
0
0
0
0
.
0
1
4
0
.
9
8
2
28
1
Fold
0
1
C
las
s
1
R
a
n
k
1
0
.
9
8
6
0
.
0
0
0
0
.
0
0
6
0
.
0
0
8
30
1
Fold
0
1
K
ap
pa
an
d
ac
c
urac
y
:
S
i
e
ga
l
an
d
Cas
t
el
l
an
(
19
9
8),
C
arle
tta
(
19
9
9),
k
ap
pa
c
on
tr
ol
ag
r
ee
me
nt
P
(
A
)
f
or
ag
r
ee
me
nt
c
h
an
g
e
P
(
e).
K
ap
pa
i
s
an
i
nte
r
-
r
ate
qu
a
l
i
t
at
i
v
e
ag
r
ee
me
nt
for
c
ate
go
r
i
c
al
i
t
em
s
a
nd
r
ob
us
t m
e
as
ure
c
al
c
u
l
at
i
on
,
w
he
r
e P
r
(
e)
i
s
c
a
l
c
ul
ati
on
,
k
ap
p
a i
s
c
orr
ec
ted
me
as
ure
c
l
as
s
i
f
i
c
ati
on
f
or
t
r
ue
c
l
as
s
es
,
p
oo
r
<
=
0.
2,
F
ai
r
i
s
eq
u
al
to
0.2
to
0.4
,
mo
d
erate
i
s
in
-
be
t
wee
n 0
.40
to
0.
60
,
go
od
ac
c
urac
y
i
s
0.
6
0 t
o
0.8
0
,
ex
c
el
l
en
t
m
ea
s
ure
i
s
0
.80
t
o 1
.
0.
K
ap
pa
s
tat
i
c
s
i
s
a
me
an
f
or
ev
a
l
ua
ti
ng
t
he
pred
i
c
ti
o
n
c
l
as
s
i
fi
er
pe
r
form
an
c
e
ac
r
os
s
al
l
i
ns
t
an
c
es
.
Cl
as
s
i
f
i
c
ati
on
me
as
ure
for
N
i
tem
s
i
n
to
mu
tua
l
l
y
ex
c
l
us
i
v
e
k
ap
pa
m
ea
s
ure
me
nt
i
s
de
s
c
r
i
be
d
in
(
1
)
:
K
=
(
Pr
(
e
)
−
Pr
(
e
)
)
/
(
(
1
−
pr
(
e
)
)
(
1)
P
r
(
a)
proba
bi
l
i
ty
of
c
l
as
s
i
fi
c
ati
on
s
uc
c
es
s
,
ac
c
urac
y
,
P
r
(
e)
probab
i
l
i
ty
to
c
h
an
c
e
of
s
uc
c
es
s
,
P
r
(
e)
r
ep
l
ac
ed
to
P
r
(
b)
ag
r
ee
me
nt
me
as
ur
e
c
l
as
s
i
f
i
c
at
i
on
a
a
nd
b,
S
u
mm
ary
o
f
s
am
p
l
e
s
i
z
es
of
22
57
,
22
5
6,
22
5
6,
22
55
,
2
25
6,
2
25
5
etc
,
by
re
-
s
am
p
l
i
ng
r
es
ul
ts
ac
r
os
s
t
un
i
ng
p
aramet
ers
wi
th
r
an
do
m
fores
t
c
l
as
s
i
f
i
er
Cr
os
s
-
V
al
i
da
t
ed
(
10
f
ol
d)
an
d
K
ap
pa
m
ea
s
ure
me
nt
are
s
ho
wn
i
n
T
ab
l
e
3.
T
ab
l
e
3.
R
an
do
m F
ores
t
Cl
as
s
i
fi
er
A
c
c
urac
y
an
d
K
ap
p
a
Me
as
urem
en
ts
of
Q
W
S
Da
ta
F
i
ne
T
un
i
ng
S
N
o
R
a
n
d
o
m
For
e
s
t
A
c
c
u
r
a
c
y
K
a
p
p
a
Mea
s
u
r
e
1
0
.
9
7
8
0
6
9
9
0
.
9
7
0
4
3
5
2
2
0
.
9
8
3
2
5
2
4
0
.
9
7
7
4
3
6
4
3
0
.
9
8
2
0
5
0
8
0
.
9
7
5
8
2
0
1
3.2.
Artif
ici
al
Neur
al
N
etw
o
r
k (
ANN)
NN
c
on
s
i
s
ts
of
an
i
n
pu
t
l
ay
er
whi
c
h
ac
c
ep
ts
i
np
ut
da
ta
he
r
e
Q
W
S
d
ata
s
et
s
ev
en
att
r
i
b
ute
v
al
ue
s
,
t
he
h
i
dd
en
l
ay
er
us
ed
f
or
we
i
g
hi
n
g
fa
c
tors
to
c
a
l
c
ul
ate
pre
di
c
ti
on
s
an
d
forw
ard
to
t
he
ou
t
pu
t
l
ay
er,
wh
i
c
h
pred
i
c
ts
th
e
c
l
as
s
l
a
be
l
s
.
A
ne
uron
i
s
c
a
l
l
ed
proc
e
s
s
i
ng
e
l
e
me
n
ts
(
A
r
ti
fi
c
i
al
n
eu
r
o
n
i
s
b
i
o
l
og
i
c
al
ne
urons
)
.
I
n
F
i
gu
r
e
3
s
h
ows
A
NN,
l
e
t
t
he
r
e
be
‘
n’
i
np
uts
{
X
i1
,
X
i2
,
…,
X
ij
}
,
h
ere
we
h
av
e
tak
en
s
ev
en
i
np
u
t
att
r
i
bu
t
es
of
Q
W
S
da
t
a,
E
ac
h
X
ij
i
s
as
s
oc
i
ate
d
wi
t
h
wei
g
hts
W
i
j
,
B
i
as
(
c
)
i
s
a
n
etwo
r
k
,us
ed
to
c
a
l
c
ul
at
e
t
h
e
ne
t
-
i
np
ut
by
ad
d
i
n
g
i
np
ut
X
ij
.,
T
hres
ho
l
d
(Ɵ
i
)
,
i
t
i
s
the
r
ea
c
he
d
to
ex
c
ee
d
th
e
v
al
ue
o
f
i
n
pu
t
ne
uro
ns
.
O
utp
ut
(
O
)
i
s
a
fte
r
ex
ec
ut
i
ng
op
erat
i
o
ns
i
s
a
no
nl
i
ne
ar
f
un
c
ti
o
na
l
(F
i
)
v
a
l
u
e
[
23
,
2
4
].
W
e
ha
v
e
c
on
du
c
te
d
ex
p
erim
e
nts
us
i
ng
ne
ura
l
ne
t
wor
k
of
Q
W
S
d
ata
s
et
2
50
7
s
a
mp
l
es
,
s
e
v
en
pred
i
c
tors
an
d
fou
r
c
l
a
s
s
es
'Cla
s
s
1',
'Cla
s
s
2', 'Cl
as
s
3', a
n
d 'c
l
as
s
4'.
T
he
g
en
er
ate
d
A
l
go
r
i
th
m
2
i
s
di
s
c
us
s
ed
be
l
ow.
A
l
g
orit
hm
2.
A
NN: N
eu
r
a
l
n
etwo
r
k
-
c
l
as
s
i
fi
c
at
i
o
n a
nd
pr
ed
i
c
t
i
on
of
c
l
as
s
l
ab
el
s
o
f Q
W
S
da
t
a
INPUT
:
D is an
input dataset of training tuples which are associated with
target values Rank1,2,3, and 4.
L:
learning rate of the network, Feedforward network use multilayers
for accuracy
OUTPUT
:
A Trained NN with testing data to classify and predict class labels
1,2,3 a
nd 4
BEGIN
Step1:
Initialize weights, bias and in networks and biases in the network
Repeat until the condition is not satisfied
BEGIN
Step2: Each training data records Xi, in D
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 1
69
3
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No.
6,
Dec
e
mb
er
20
1
9
:
31
9
1
-
3202
3196
BEGIN
Step3:
// Propagate input data forward to layer
Step4:
For input of
each layer with j
BEGIN
Compute Result (Fi) = Input (Xi);// result of i/p is actual i/p
data.
//if More than one hidden layer improves the performance accuracy
S
tep5: each hidden layer or result at output layer unit J
NetInput (Xj)=∑WijFij+Ɵj // calcula
te net input j which
corresponds to the previous layer i, and compute Function output
Fi as Output (Oj)= (1/1+e
-
lj
), calculate result output(sigmoid)
for input each j value.
each input j
END
// Step 4 end
Step6: propagation errors if any
each unit for j in the results of the output layer
Errj=Fj(1
-
Fj)(Tj
-
Fj);// compute the error if any
For each input j in hidden layers; from considering last to the
first hidden layer
Compute Errj=Fj(1
-
Fj) ∑k Errk
Wjk
;//
calculate error respect to
next higher layer k
weight for W
ij
in the network associated with each layer and
value
BEGIN
ΔWij=1+Err j Fj// increment of weight
Wij= Wij+ ΔWij;// update the weight value
END
Step7:
For bias Ɵ in
each network update value
Δ Ɵj=(l) + Err j // Bias increment
END
END
END
F
i
gu
r
e
3.
A
r
ti
f
i
c
i
a
l
N
e
ural
N
etwo
r
k
M
o
de
l
T
he
ex
pe
r
i
me
nt
al
r
es
u
l
ts
c
on
du
c
te
d
us
i
ng
R
pro
gra
mm
i
ng
wi
th
Q
W
S
d
ata
a
n
d
A
NN
c
on
fus
i
o
n
ma
tr
i
x
i
s
de
s
c
r
i
b
ed
i
n
T
a
bl
e
4
.
A
NN
pre
di
c
ted
v
al
u
es
wi
th
o
bs
erv
ati
o
n
c
om
pa
r
i
s
o
n
de
s
c
r
i
be
d
i
n
T
a
bl
e
5
.
Re
-
s
am
p
l
i
ng
of
A
NN
w
i
th
K
a
pp
a
me
as
ur
es
and
t
he
r
es
ul
t
s
are
de
s
c
r
i
b
ed
i
n
T
a
bl
e
6.
T
ab
l
e
4.
A
r
ti
f
i
c
i
a
l
N
eu
r
a
l
Ne
twork
Con
f
us
i
o
n
Ma
tr
i
x
P
r
e
d
ic
t
ion
R
a
n
k
1
R
a
n
k
2
R
a
n
k
3
R
a
n
k
4
C
las
s
-
1
587
1
2
1
C
las
s
-
2
2
803
2
3
C
las
s
-
3
2
1
587
4
C
las
s
-
4
2
1
8
501
A
c
c
u
r
a
c
y
(
a
v
e
r
a
g
e
)
:
9
8
.
1
1
%
The
A
c
c
u
r
a
c
y
(
a
v
e
r
a
g
e
)
:
9
8
.
1
1
%
f
ine
-
t
u
n
ing
a
c
c
u
r
a
c
y
is
9
8
.
2
0
A
c
c
u
r
a
c
y
=9
,
a
n
d
d
e
c
a
y
i
s
9
a
n
d
0
.
0
6
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Cl
as
s
i
f
i
c
ati
on
o
f we
b s
erv
i
c
es
us
i
ng
da
t
a m
i
n
i
ng
al
go
r
i
t
hm
s
... (
M.
S
w
am
i
Das
)
3197
T
ab
l
e
5.
A
r
ti
f
i
c
i
a
l
N
eu
r
a
l
Ne
twork
Con
f
us
i
o
n
Ma
tr
i
x
S
N
o
P
r
e
d
ic
t
ion
Ob
s
e
r
v
a
t
ion
R
a
n
k
1
R
a
n
k
2
R
a
n
k
3
R
a
n
k
4
R
o
w
I
n
d
e
x
S
iz
e
1
C
las
s
-
2
R
a
n
k
2
1
.
3
8
2
8
e
-
01
8
.
0
4
4
0
e
-
01
1
.
6
9
4
6
e
-
04
5
.
7
1
3
8
e
-
02
7
1
2
C
las
s
1
R
a
n
k
4
5
.
6
4
6
6
e
-
01
5
.
7
6
6
2
e
-
02
5
.
7
4
2
1
e
-
03
3
.
7
1
9
2
e
-
01
13
1
3
C
las
s
-
2
R
a
n
k
2
5
.
3
7
0
0
e
-
02
9
.
2
6
6
8
e
-
01
3
.
7
2
4
5
e
-
05
1
.
9
5
7
2
e
-
02
18
1
4
C
las
s
2
R
a
n
k
2
6
.
1
4
6
7
e
-
02
9
.
1
5
6
9
e
-
01
4
.
6
0
4
3
e
-
05
2
.
2
7
8
6
e
-
02
25
1
5
C
las
s
4
R
a
n
k
4
4
.
7
5
0
7
e
-
01
1
.
0
3
5
6
e
-
03
3
.
6
1
8
7
e
-
02
4
.
8
7
6
9
e
-
01
34
1
T
ab
l
e
6.
R
e
-
s
am
p
l
i
ng
R
es
u
l
ts
ac
r
os
s
T
un
i
ng
P
ara
me
te
r
s
us
i
ng
A
NN w
i
th
K
a
pp
a
M
ea
s
ures
S
iz
e
D
e
c
a
y
A
c
c
u
r
a
c
y
K
a
p
p
a
1
0
.
0
5
0
.
9
7
4
0
7
3
2
0
.
9
6
5
0
7
3
8
2
0
.
0
6
0
.
9
8
5
2
0
6
0
0
.
9
7
5
8
3
7
8
3
0
.
0
2
0
.
9
7
7
6
5
7
3
0
.
9
6
9
8
9
9
1
4
0
.
0
1
0
.
9
7
6
8
7
7
9
0
.
9
6
8
8
3
5
2
5
0
.
0
1
0
.
9
7
9
2
6
3
6
0
.
9
7
2
0
6
7
5
3.3
.
J4
8 D
ec
i
sion
Tre
e
Is
an
ap
pro
ac
h
whi
c
h
ac
ts
as
a
c
l
as
s
i
fi
c
at
i
on
pr
ed
i
c
t
or
fr
om
the
l
i
s
t
of
v
al
ue
s
,
the
target
i
s
de
p
en
da
b
l
e v
a
l
u
e,
a
nd
i
t
i
s
us
ed
to
predi
c
t
ta
r
g
et
v
al
ue
.
T
h
e J
48
Dec
i
s
i
on
Tr
e
e
i
s
de
s
c
r
i
be
d
i
n
A
l
g
or
i
t
hm
3.
A
l
g
orit
hm
3.
J
48
-
C
l
as
s
i
f
i
c
at
i
on
of
Q
W
S
da
ta
us
i
n
g D
ec
i
s
i
on
Tr
ee
s
DTreeQoS(
Datapartition
Dp,
data attributes
)
Input: QWS data
Attributelist
(RT,AV,TH,SUCC,REL,COMP,BP)
RT
Response
time, AV: Availability, TH: Throughput, SUCC, Successability, REL:
Reliability, COM: Compliance,
BP:
Best practices.
-
QWS
selectionMethod: Procedure for
splitting into
individual classes.
Output: Decision Tree with class labels class
-
1,2,3 and 4.
Step1.
Create a new node N
Step2.
If data records partition Dp of the same class of type(1,2,3, and 4),
then return the leaf
Nod
e N label with relevant class type as 1,2,3 and 4.
Step3.
If (data attribute list)is EMPTY thenReturn N, leaf node, the
majority of classes in data partition Dp / // majority of classes
of class type
Step4
.
apply QWSselectionMethod(Dp,
data attribute
list
) to find the
best
splitting criterion.
Step5
.
Splitting criterion label N
Step6
.
If
splitting data values
and
Multisplit
permitted then//
Step7
.
Adddataattributelist
data attribute list
–
splitting attribute
; //
removing splitting
attribute
Step8
.
For the
result
of the splitting criterion of each // To find classify
label
Step9
.
Let
Dresult
be set of data records in Da satisfy result; //
apartition result
Step10
.
If
Dresult
is
EMPTY
then add to leaf label of majority class label
i
n Da to node N; else add a node returned by DTreeQoS (
Dresult
,
dataattributelist
) to node N;
Step12
.
end for
Step13
.
Return N
A
pp
l
y
i
n
g
A
l
g
orit
h
m3
(
J
48
Dec
i
s
i
o
n
tr
ee
)
o
n
Q
W
S
d
a
ta,
wh
i
c
h
c
on
s
i
s
ts
of
25
07
r
ec
ord
s
am
pl
es
,
r
es
u
l
ts
i
nto
fo
ur
c
l
as
s
es
na
me
l
y
:
'C
l
as
s
1',
'
Cl
as
s
2',
'Cl
as
s
3',
a
nd
'Cl
as
s
4'
us
i
n
g
s
ev
e
n
predi
c
tors
.
Re
-
s
am
pl
i
ng
:
C
r
os
s
-
V
al
i
da
te
d
(
1
0
f
ol
d),
u
s
i
ng
R
l
an
gu
a
ge
the
s
um
ma
r
y
of
s
am
p
l
e
s
i
z
es
of
22
5
6,
22
56
,
2
25
7
,
22
56
,
22
5
7,
2
25
6,
etc
.
the
r
es
ul
ts
s
ho
wn
i
n
T
ab
l
e
7,
a
nd
a
c
c
urac
y
i
s
97
.5
6
63
7%
an
d
K
ap
pa
v
al
ue
i
s
9
6.7
22
4
8%.
T
un
i
ng
w
i
th
pa
r
a
me
t
er
'C'
was
a
c
o
ns
tan
t
v
al
ue
of
0.2
5.
Tun
i
ng
pa
r
a
me
t
er at
c
on
s
tan
t v
a
l
ue
M
i
s
he
l
d
at
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 1
69
3
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No.
6,
Dec
e
mb
er
20
1
9
:
31
9
1
-
3202
3198
T
ab
l
e
7.
T
he
C
om
p
arati
v
e
S
tud
y
of
J
45
Cl
as
s
i
fi
c
at
i
on
P
r
ed
i
c
ti
o
ns
wi
t
h Obs
erv
ati
o
ns
S
N
o
P
r
e
d
ic
t
ion
Ob
s
e
r
v
a
t
ion
R
a
n
k
1
R
a
n
k
2
R
a
n
k
3
R
a
n
k
4
R
o
w
I
n
d
e
x
S
iz
e
1
C
las
s
2
R
a
n
k
2
0
.
0
0
0
0
1
.
0
0
0
0
0
.
0
0
0
0
0
.
0
0
0
0
2
0
.
2
5
2
C
las
s
1
R
a
n
k
1
0
.
9
9
7
9
0
.
0
0
2
0
0
.
0
0
0
0
0
.
0
0
0
0
5
0
.
2
5
3
C
las
s
1
R
a
n
k
1
0
.
9
9
7
9
0
.
0
0
2
0
0
.
0
0
0
0
0
.
0
0
0
0
10
0
.
2
5
4
C
las
s
1
R
a
n
k
1
0
.
9
9
7
9
0
.
0
0
2
0
0
.
0
0
0
0
0
.
0
0
0
0
12
0
.
2
5
5
C
las
s
4
R
a
n
k
4
0
.
0
0
0
0
0
.
0
0
0
0
0
.
0
0
0
0
1
.
0
0
0
0
16
0
.
2
5
3.4
.
eX
t
r
e
me G
r
adient
Bo
o
sti
n
g
(
X
G
b
o
o
st T
ec
h
n
iqu
e)
T
hi
s
i
s
a
ma
c
h
i
n
e
l
e
arni
ng
al
go
r
i
th
m,
s
up
erv
i
s
e
d
l
e
ar
ni
n
g
to
f
i
nd
t
he
tas
k
s
r
eg
r
es
s
i
on
,
r
an
k
i
ng
,
an
d
c
l
as
s
i
fi
c
at
i
o
n.
P
r
ed
i
c
ti
o
n
(
Y
i
)
for
gi
v
e
n
a
v
al
u
e
(
X
i
,)
us
i
n
g
a
l
i
n
ea
r
mo
de
l
i
s
us
ed
to
predi
c
t
i
on
Y
'
i
=
∑Ɵ
i
X
i
,
w
he
r
e
wei
gh
t
i
np
ut,
t
he
pred
i
c
ti
on
,
c
a
n
h
av
e
d
i
ffe
r
e
nt
i
nte
r
pret
ati
on
s
,
the
tas
k
de
p
en
ds
on
r
eg
r
e
s
s
i
on
or
c
l
as
s
i
fi
c
a
ti
o
n,
Ɵ
i
co
-
eff
i
c
i
en
t
de
n
ote
p
aram
et
er
the
mo
de
l
i
s
us
ed
f
or
r
an
k
i
ng
th
e
o
utp
uts
.
X
G
bo
os
t
tec
hn
i
q
ue
(
eX
tr
e
me
grad
i
en
t
b
oo
s
ti
n
g
)
us
es
s
ev
e
n
predi
c
tor
s
,
fo
ur
c
l
as
s
es
as
Cl
as
s
-
1,2
,3
an
d
4,
ex
p
erim
e
nts
at
Cr
os
s
-
V
al
i
d
at
ed
(
10
fo
l
d)
of
s
am
pl
e
s
i
z
es
:
2
25
8,
2
25
5
,
22
56
,
2
25
6,
22
5
7,
22
5
7
,
et
c
.
T
he
ex
pe
r
i
me
nta
t
i
o
ns
w
i
t
h
X
G
B
o
os
ti
n
g
me
th
od
r
es
u
l
ts
are
s
ho
w
n
i
n
T
ab
l
e
8,
an
d
T
ab
l
e
9
X
G
bo
os
t
c
on
f
us
i
o
n
ma
tr
i
x
ac
r
os
s
tun
i
ng
pa
r
am
ete
r
s
.
T
ab
l
e
8.
eX
tr
e
me
G
r
a
di
en
t
B
oo
s
t
i
ng
Cl
as
s
i
f
i
c
ati
on
o
f Q
W
S
C
om
pa
r
i
s
on
s
A
c
c
urac
y
wi
th
K
a
pp
a
Me
as
ures
S
N
o
The
t
a
A
lph
a
N
r
o
u
n
d
s
A
c
c
u
r
a
c
y
1
0
e
+0
0
0
e
+0
0
50
0
.
9
8
2
8
3
9
7
2
0
e
+0
0
0
e
+0
0
100
0
.
9
8
2
8
3
9
7
3
0
e
+0
0
0
e
+0
0
150
0
.
9
8
2
8
3
9
7
4
0
e
+0
0
1e
-
04
50
0
.
9
8
3
2
3
8
1
5
1e
-
01
1e
-
04
100
0
.
9
8
3
6
0
4
6
T
ab
l
e
9.
eX
tr
e
me
G
r
a
di
en
t
B
oo
s
t
i
ng
Conf
us
i
on
Ma
tr
i
x
P
r
e
d
ic
t
ion
R
a
n
k
1
R
a
n
k
2
R
a
n
k
3
R
a
n
k
4
C
las
s
-
1
583
1
0
1
C
las
s
-
2
3
800
0
3
C
las
s
-
3
0
2
590
10
C
las
s
-
4
7
3
9
495
A
c
c
u
r
a
c
y
(
a
v
e
r
a
g
e
)
:
9
8
.
4
4
%
The
A
c
c
u
r
a
c
y
(
a
v
e
r
a
g
e
)
:
9
8
.
4
4
%
f
ine
t
u
n
ing
a
c
c
u
r
a
c
y
is
9
8
.
4
5
%
c
o
n
s
t
a
n
t
v
a
lue
is
0
.
3
o
p
t
im
u
m
mod
e
l
lar
g
e
s
t
v
a
lue
,
5
0
r
o
u
n
d
s
lamda
=0
.
1
,
a
c
c
u
r
a
c
y
=9
,
a
n
d
d
e
c
a
y
i
s
9
a
n
d
0
.
0
6
3.5
.
K
-
Near
es
t
Ne
igh
b
o
r
s
Cl
a
ss
if
ica
t
ion
(
KN
N)
K
NN
c
l
as
s
i
f
i
c
at
i
on
a
l
g
orit
h
m
[25
]
i
np
ut
ha
v
e
‘
k
’
c
l
os
et
tr
ai
n
i
n
g
tup
l
es
i
n
the
f
ea
tur
e
s
el
ec
ti
on
,
the
o
bj
ec
t
b
ei
n
g
as
s
i
gn
e
d
t
o
t
he
mo
s
t
c
om
mo
n
i
s
‘
k
’
ne
ares
t
n
ei
gh
b
ors
.
F
i
nd
th
e
mo
s
t
s
i
mi
l
ar
o
bj
ec
t
fr
om
tr
a
i
n
i
ng
da
ta;
the
tes
ti
ng
d
ata
w
i
l
l
l
o
ok
i
nto
the
tr
ai
ni
n
g
d
ata
s
et
to
be
the
mo
s
t
s
i
mi
l
ar
ob
j
ec
t
b
as
ed
o
n
fea
ture
s
el
ec
t
i
on
an
d
di
s
ta
nc
e
fun
c
ti
o
ns
.
K
N
N
al
go
r
i
t
hm
i
s
de
s
c
r
i
be
d
i
n
A
l
g
orit
hm
4.
A
l
g
orit
hm
4.
K
NN
A
pp
r
o
ac
h
Input: Given records and attributes S, from Matrix a=[Aij]
Output:
classification label (class
-
1,2,3 and 4)
Each record classified into Rank the web service (
WsRF:
Web service Relevancy
Function)
begin
Step1.
Given a set of records and attributes, form matrix A= [a
ij
]
Step2.
For each record P in the Test data
d
o
For each record in A, in the training data
A do
Calculate the similarity of input testing data which is most similar
to
trained
dataset
EUDistance
(P,A)
Store DIS_Array(P
i
,A
i
) and find the class label.
end for
end for
end
Q
W
S
d
ata
s
et
25
0
7
s
am
p
l
es
,
for
s
ev
en
pre
di
c
tors
an
d
i
nto
f
ou
r
c
l
as
s
es
:
'R
an
k
1',
'Rank
2',
'R
an
k
3',
'Rank
4',
u
s
e
of
c
r
os
s
-
v
al
i
da
t
i
on
,
1
0
fo
l
d
t
he
s
am
p
l
e
s
i
z
es
of
22
55
,
2
2
56
,
22
56
,
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
MNIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Cl
as
s
i
f
i
c
ati
on
o
f we
b s
erv
i
c
es
us
i
ng
da
t
a m
i
n
i
ng
al
go
r
i
t
hm
s
... (
M.
S
w
am
i
Das
)
3199
22
57
,
22
58
,
2
25
7
etc
,
an
d
Re
-
s
am
p
l
i
n
g
r
es
ul
ts
ac
r
os
s
tun
i
ng
p
arame
ters
an
d
r
es
ul
t
c
on
du
c
te
d
s
ho
wn
i
n
T
a
bl
e 1
0
. K
N
N w
i
th
k
v
al
u
e 5
i
s
ma
x
i
mu
m a
c
c
urac
y
of
89
.3
4%
.
3.6
.
S
V
M wit
h
Radia
l Ba
sis
Fu
n
ctio
n
Ker
n
e
l
Ap
p
r
o
a
ch
T
hi
s
i
s
a
m
eth
od
us
e
d
to
c
l
as
s
i
fy
the
d
ata
.
S
V
M
c
l
as
s
i
fi
c
ati
on
us
e
ma
x
i
mi
z
e
ma
r
gi
n
f
or
ac
c
urate
v
al
ue
s
.
It
i
s
a
l
i
n
ea
r
c
l
as
s
i
f
i
er,
w
h
ere
A
i
i
s
i
np
ut,
W
i
s
v
aria
b
l
e
o
f
a
s
tr
ai
g
ht
l
i
n
e
wi
t
h
c
on
s
tan
t
B
.
IN
(
2
)
-
(
5
)
de
p
i
c
ts
th
e p
r
oc
es
s
i
ng
:
bF
(
Ai
,
W
,
B
)
=
s
ign
(
A
iW
+
B
)
(
2)
where
A
i
,
i
s
i
np
u
t d
a
ta
w
hi
c
h i
s
a
v
ar
i
ab
l
e
an
d
B
i
s
c
on
s
tan
t o
f
a s
tr
ai
gh
t
l
i
ne
:
W
A
i
+
b
≥
1
(
3)
i
f
F
i
=
+
1
w
he
r
e
W
i
s
a
v
a
r
i
ab
l
e
of
l
i
ne
,
ts
he
l
i
ne
ar
l
i
ne
W
A
i
+
b
i
s
ma
x
i
m
i
z
ed
m
argi
n
M=
2/
|
W
|
,
mi
n
i
m
i
z
e 1
/
2W
t
w w
i
th
t
he
s
ub
j
ec
t to
Mi
n
i
m
i
z
e (w)
=
1/2
W
t
W
an
d s
ub
j
ec
t t
o o
utp
u
t
(
wA
i
+
b) ≥
1
W
A
i
+
b
≤
1
(
4)
i
f Fi
=
-
1
where
W
i
s
a
v
ari
a
bl
e
an
d
(
+
)
≥
1
(
5)
f
or al
l
r
em
ai
ni
ng
v
a
l
ue
s
.
Q
W
S
da
t
as
et
2
50
7
s
am
p
l
e
s
ex
pe
r
i
m
en
ts
c
on
du
c
te
d
u
s
i
ng
R
progr
a
mm
i
n
g
w
i
th
s
ev
en
predi
c
tors
a
nd
f
ou
r
c
l
as
s
es
:
'Rank
1',
'R
an
k
2',
'R
an
k
3',
'Rank
4',
r
e
-
s
am
pl
i
ng
,
s
am
p
l
i
n
g
th
e
us
e
o
f
c
r
os
s
-
v
al
i
da
ti
on
ten
-
f
ol
d
,
t
he
r
es
u
l
ts
us
i
ng
S
V
M
wi
t
h
R
B
F
K
me
t
ho
d
are
s
ho
w
n
i
n
T
a
bl
e
1
1.
S
um
ma
r
y
of
s
a
mp
l
e
s
i
z
es
:
22
57
,
22
56
,
22
5
6,
2
25
8,
2
25
6,
22
5
7,
etc
.
r
e
-
s
a
mp
l
i
n
g
r
es
ul
ts
ac
r
os
s
tun
i
ng
pa
r
am
ete
r
s
.
T
ab
l
e
10
.
The
K
NN
Cl
as
s
i
fi
c
ati
on
of
Q
W
S
Data
an
d
Com
pa
r
i
s
o
ns
wi
t
h A
c
c
urac
y
wi
t
h
K
ap
pa
Me
as
ure
me
nt
S
N
o
K
v
a
lue
A
c
c
u
r
a
c
y
K
a
p
p
a
1
5
0
.
8
9
3
4
7
9
4
0
.
8
5
5
6
5
9
6
2
7
0
.
8
8
5
5
0
9
6
0
.
8
4
4
6
6
7
7
3
9
0
.
8
7
3
9
3
4
9
0
.
8
2
8
7
7
8
2
A
c
c
u
r
a
c
y
i
s
a
t
k
=5
,
o
p
t
im
a
l
mod
e
l
,
t
h
e
lar
g
e
s
t
v
a
lue
T
ab
l
e
11
.
S
V
M
wi
t
h Ra
di
a
l
B
as
i
s
Fu
nc
ti
o
n K
erne
l
wi
th
c
l
as
s
i
fy
Q
W
S
Da
ta
A
c
c
urac
y
an
d K
ap
p
a
Me
as
ure
me
n
ts
S
N
o
C
S
igma
A
c
c
u
r
a
c
y
K
a
p
p
a
1
1
0
.
1
0
.
9
3
9
7
4
9
4
0
.
9
1
8
4
6
2
0
2
1
0
.
2
0
.
9
4
8
9
0
3
2
0
.
9
3
0
8
7
7
3
1
0
.
3
0
.
9
5
4
8
9
3
7
0
.
9
3
9
0
3
8
7
4
1
0
.
4
0
.
9
5
4
8
9
8
5
0
.
9
3
9
0
5
9
3
5
1
0
.
5
0
.
9
5
7
3
0
1
8
0
.
9
4
2
3
3
1
4
A
c
c
u
r
a
c
y
a
t
lar
g
e
s
t
o
p
t
i
mal
v
a
lue
w
h
e
n
s
igma
=
0
.
6
a
n
d
C
=
1
4.
Result
s
and
Dis
cussion
s
W
S
ob
j
ec
ts
n
ee
d
th
e
q
ua
l
i
t
y
of
s
erv
i
c
es
;
t
hi
s
i
s
as
s
oc
i
ate
d
wi
t
h
e
ac
h
ob
j
ec
t,
i
n
bu
s
i
ne
s
s
proc
es
s
c
oo
r
di
n
ati
ng
t
o
de
al
w
i
th
s
erv
i
c
es
a
nd
ma
n
a
gi
n
g
s
erv
i
c
e
qu
al
i
t
i
es
.
T
o
prov
i
de
qu
al
i
ty
s
erv
i
c
es
ac
c
ordi
ng
to
us
er
r
eq
ui
r
e
me
nts
,
W
S
r
ea
l
-
ti
m
e
ap
p
l
i
c
at
i
o
ns
us
e
of
v
i
de
o,
tex
t,
i
m
ag
e
,
an
d
o
the
r
e
l
e
me
nts
by
us
er
s
,
Q
oS
of
w
eb
s
erv
i
c
e,
s
at
i
s
fi
es
the
e
nd
-
us
er
r
eq
ue
s
ts
.
F
or
ex
am
p
l
e,
the
r
es
p
on
s
e
ti
me
of
W
e
bs
i
te
qu
al
i
ty
i
s
as
s
oc
i
ate
d
wi
t
h
v
ari
ou
s
p
aram
ete
r
s
l
i
k
e
n
etwo
r
k
,
ap
p
l
i
c
at
i
on
s
e
r
v
i
c
es
.
T
h
e
de
s
i
g
ne
r
a
i
m
i
s
t
o
m
o
ni
t
or
the
q
ua
l
i
ty
c
on
t
en
ts
an
d
en
s
ure
qu
al
i
ty
s
erv
i
c
es
.
T
he
c
us
to
me
r
s
r
eq
u
i
r
e
q
ua
l
i
ty
s
erv
i
c
es
,
hi
g
h
av
a
i
l
a
bi
l
i
ty
,
s
ec
urit
y
,
c
os
t
op
t
i
m
i
z
at
i
on
,
and
oth
ers
[
26
].
W
S
l
ay
er,
qu
a
l
i
ty
m
on
i
tore
d,
a
nd
ad
j
u
s
ted
pa
r
a
me
t
ers
,
for
ex
a
m
pl
e
ba
n
dwi
dth
,
c
om
mu
ni
c
at
i
on
l
ay
er
by
w
eb
s
erv
i
c
e,
wh
i
c
h
ha
nd
l
es
me
s
s
ag
e
c
on
t
en
ts
i
n
a
r
ea
l
-
t
i
me
l
ay
er
whi
c
h
c
o
mm
u
ni
c
at
es
wi
t
h
s
erv
i
c
es
be
twe
en
c
l
i
en
t
an
d
s
erv
er.
T
h
e
we
b
s
erv
i
c
es
qu
a
l
i
ty
pa
r
am
ete
r
s
of
a
c
c
es
s
c
on
t
r
ol
the
i
nf
ormat
i
on
of
da
ta
au
d
i
o,
v
i
d
eo
,
a
tex
t
do
c
u
m
en
t
an
d
oth
er
do
c
um
e
nts
w
i
th
v
ari
ou
s
p
aramet
ers
i
nf
l
ue
nc
e
to
me
as
ure
the
qu
al
i
ty
of
s
o
ftware.
Her
e
we
ha
v
e
tak
en
Q
W
S
da
t
as
et,
an
d
the
r
es
u
l
ts
are
ex
ec
ut
ed
us
i
ng
R
-
La
ng
ua
g
e
a
nd
i
de
nt
i
fi
ed
mo
s
t
i
mp
ortant
i
nf
l
ue
nc
e
pa
r
a
m
ete
r
s
are
be
s
t
prac
ti
c
es
,
s
uc
c
es
s
ab
i
l
i
ty
,
av
ai
l
a
bi
l
i
ty
,
r
es
po
ns
e
t
i
me
,
r
el
i
a
bi
l
i
ty
,
thro
ug
h
pu
t,
a
nd
c
om
p
l
i
an
c
e
s
ho
wn
i
n
F
i
gu
r
e
4
i
n
web
ap
pl
i
c
at
i
on
de
v
e
l
op
me
nt
.
T
he
ex
i
s
t
i
ng
web
s
erv
i
c
e c
l
as
s
i
fi
c
at
i
on
me
t
ho
ds
an
d
i
t
s
ac
c
urac
y
v
al
ue
s
are s
ho
wn i
n Tab
l
e
12
.
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 1
69
3
-
6
93
0
T
E
L
KO
MNIK
A
V
ol
.
17
,
No.
6,
Dec
e
mb
er
20
1
9
:
31
9
1
-
3202
3200
T
ab
l
e
12
.
W
eb
S
erv
i
c
e C
l
a
s
s
i
fi
c
ati
on
s
of
E
x
i
s
t
i
ng
Me
t
ho
ds
a
nd
A
c
c
urac
y
V
al
ue
s
S
N
o
C
las
s
i
f
i
c
a
t
ion
Me
t
h
o
d
A
c
c
u
r
a
c
y
1
N
a
ï
v
e
b
a
y
s
8
3
.
6
2
%
2
Gr
o
u
p
Met
h
o
d
o
f
D
a
t
a
H
a
n
d
li
n
g
(
G
MDH
)
9
8
.
3
2
%
3
B
a
c
k
p
r
o
p
a
g
a
t
ion
a
n
d
N
e
u
r
a
l
N
e
t
w
o
r
k
s
(
B
P
N
N
)
9
7
.
2
2
%
T
he
ac
c
urac
y
v
al
u
es
of
c
l
as
s
i
fi
c
at
i
on
me
t
ho
ds
i
mp
l
e
me
nt
ed
us
i
ng
R
La
ng
u
ag
e,
an
d
r
es
ul
ts
are
s
ho
w
n
i
n
F
i
g
u
r
e
5,
c
om
p
ute
d
c
l
as
s
i
f
i
c
at
i
on
ac
c
urac
y
,
an
d
k
ap
p
a
me
as
ure
me
n
ts
graph
are
a
l
s
o
s
ho
w
n.
F
i
gu
r
e
6
s
h
ows
th
e
r
es
ul
ts
W
S
c
l
as
s
i
fi
c
a
ti
o
ns
us
i
ng
(
A
NN)
m
eth
o
d.
T
he
F
i
gu
r
e
7
s
ho
ws
th
e
ac
c
urac
y
of
r
a
nd
o
ml
y
s
el
ec
te
d p
r
e
di
c
tors
.
F
i
g
ure
8
s
ho
w
s
th
e
ac
c
urac
y
,
nu
mb
er
of
i
terat
i
on
s
w
eb
s
erv
i
c
e
c
l
as
s
i
fi
c
at
i
on
s
b
y
ex
tr
em
e
G
r
ad
i
en
t
bo
os
ti
ng
me
t
ho
d.
Rand
om
l
y
s
el
ec
t
ed
pred
i
c
t
ors
for
i
s
s
h
own
i
n
F
i
gu
r
e
7,
i
n
w
hi
c
h
th
e
pre
di
c
t
i
on
at
a
m
i
ni
mu
m
ac
c
urac
y
by
K
NN i
s
89
.3
9
% and
ma
x
i
mu
m
ac
c
urac
y
by
X
G
bo
os
t te
c
hn
i
q
ue
s
are
98
.4
4%.
F
i
gu
r
e
4.
I
mp
ort
an
c
e o
f
W
e
b
S
erv
i
c
es
F
i
gu
r
e
5.
W
eb
S
erv
i
c
e
C
l
a
s
s
i
fi
c
ati
on
ac
c
urac
y
wi
th
k
ap
pa
me
as
ureme
nts
F
i
gu
r
e
6.
W
eb
s
erv
i
c
e c
l
as
s
i
fi
c
at
i
on
us
i
n
g
A
r
ti
f
i
c
i
a
l
N
e
ural
N
etw
ork
s
F
i
gu
r
e
7.
R
an
d
om
l
y
s
el
ec
te
d p
r
ed
i
c
tors
T
he
e
X
tr
e
me
gradi
en
t
bo
o
s
ti
ng
me
t
ho
d
r
es
ul
ts
are
d
es
c
r
i
be
d
i
n
F
i
g
ure
8
tha
t
s
ho
ws
the
ac
c
urac
y
by
us
i
ng
al
ph
a
at
i
tera
ti
o
ns
wi
th
t
he
mi
ni
mu
m
whi
c
h
i
s
9
8.4
4%
a
nd
the
ma
x
i
m
um
v
al
ue
i
s
98
.
45
%
by
i
tera
ti
on
s
w
i
th
al
p
ha
v
a
l
ue
s
.
eX
tr
e
me
gra
di
en
t
bo
os
t
i
n
g
me
t
ho
d
go
t
the
hi
gh
es
t
ac
c
urac
y
c
om
pa
r
ed
w
i
th
v
ari
o
us
da
ta
mi
n
i
n
g
me
th
od
s
ac
c
urac
i
e
s
are
s
ho
wn
i
n
T
ab
l
e
13
.
T
he
ex
pe
r
i
me
n
tal
r
es
ul
ts
c
on
d
uc
ted
wi
t
h
Q
W
S
da
t
as
et
us
i
ng
R
l
a
ng
ua
ge
a
nd
v
ari
ou
s
c
l
as
s
i
fi
c
at
i
on
m
eth
o
ds
a
nd
c
om
pa
r
at
i
v
e
s
tu
dy
of
ac
c
ur
ac
y
wi
th
fi
ne
t
un
i
ng
v
al
ue
s
are
s
h
own
i
n
T
ab
l
e
13
.
Evaluation Warning : The document was created with Spire.PDF for Python.