T
E
L
K
O
M
N
I
K
A
T
elec
o
m
m
un
ica
t
io
n,
Co
m
pu
t
ing
,
E
lect
ro
nics
a
nd
Co
ntr
o
l
Vo
l.
19
,
No
.
4
,
A
u
g
u
s
t
2
0
2
1
,
p
p
.
1
2
0
8
~
1
2
1
7
I
SS
N:
1
6
9
3
-
6
9
3
0
,
ac
cr
ed
ited
First Gr
ad
e
b
y
Kem
en
r
is
tek
d
i
k
ti,
Dec
r
ee
No
: 2
1
/E/KPT
/2
0
1
8
DOI
: 1
0
.
1
2
9
2
8
/TE
L
KOM
NI
K
A.
v
1
9
i4
.
1
8
8
7
8
1208
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//jo
u
r
n
a
l.u
a
d
.
a
c.
id
/in
d
ex
.
p
h
p
/TELK
OM
N
I
K
A
Neura
l ne
twork
t
echnique
wit
h
dee
p struc
ture
f
o
r im
pro
v
ing
a
uthor ho
mo
ny
m a
nd sy
no
ny
m cla
ss
ificatio
n in
dig
ital libra
ries
F
irda
us
1
,
Siti
Nurm
a
ini
2
,
Va
rindo
O
ck
t
a
K
eneddi P
utr
a
3
,
Anni
s
a
Da
rm
a
wa
hy
un
i
4
,
Rez
a
F
irsa
nd
a
y
a
M
a
lik
5
,
M
uh
a
m
m
a
d Na
ufa
l
Ra
chm
a
t
ull
a
h
6
,
Andre
H
er
v
ia
nt
J
uli
a
no
7
,
T
io
Art
ha
Nug
ra
ha
8
1,
2,
3,
4
,
6,
7,
8
In
telli
g
e
n
t
S
y
ste
m
s Re
se
a
rc
h
G
ro
u
p
,
Un
i
v
e
rsitas
S
riwij
a
y
a
,
P
a
lem
b
a
n
g
,
In
d
o
n
e
sia
5
Co
m
m
u
n
ica
ti
o
n
Ne
two
rk
s
a
n
d
I
n
fo
rm
a
ti
o
n
S
e
c
u
ri
ty
Re
se
a
rc
h
La
b
,
Un
i
v
e
rsitas
S
riwij
a
y
a
,
P
a
lem
b
a
n
g
,
In
d
o
n
e
sia
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
3
0
,
2
0
2
0
R
ev
is
ed
J
an
3
,
2
0
2
1
Acc
ep
ted
J
an
2
0
,
2
0
2
1
Au
th
o
r
n
a
m
e
d
isa
m
b
i
g
u
a
ti
o
n
(AN
D),
a
lso
re
c
o
g
n
ize
d
a
s
n
a
m
e
-
id
e
n
ti
fica
ti
o
n
,
h
a
s
lo
n
g
b
e
e
n
se
e
n
a
s
a
c
h
a
ll
e
n
g
i
n
g
issu
e
in
b
ib
li
o
g
ra
p
h
ic
d
a
ta.
In
o
t
h
e
r
wo
rd
s,
th
e
sa
m
e
a
u
th
o
r
m
a
y
a
p
p
e
a
r
u
n
d
e
r
se
p
a
ra
te
n
a
m
e
s,
sy
n
o
n
y
m
s,
o
r
d
isti
n
c
t
a
u
th
o
rs
m
a
y
h
a
v
e
sim
il
a
r
to
th
o
se
re
fe
rre
d
to
a
s
h
o
m
o
n
y
m
s.
S
o
m
e
p
re
v
io
u
s
re
se
a
rc
h
h
a
s
p
r
o
p
o
se
d
AN
D p
ro
b
l
e
m
.
To
th
e
b
e
st o
f
o
u
r
k
n
o
wle
d
g
e
,
n
o
stu
d
y
d
isc
u
ss
e
d
sp
e
c
ifi
c
a
ll
y
s
y
n
o
n
y
m
a
n
d
h
o
m
o
n
y
m
,
w
h
e
re
a
s
su
c
h
c
a
se
s
a
re
th
e
c
o
re
in
AN
D
to
p
ic.
T
h
is
p
a
p
e
r
p
re
se
n
ts
th
e
c
las
sifica
ti
o
n
o
f
n
o
n
-
hom
o
n
y
m
-
sy
n
o
n
y
m
,
h
o
m
o
n
y
m
-
sy
n
o
n
y
m
,
s
y
n
o
n
y
m
,
a
n
d
h
o
m
o
n
y
m
c
a
se
s
b
y
u
sin
g
t
h
e
DBLP
c
o
m
p
u
ter
sc
ien
c
e
b
ib
li
o
g
r
a
p
h
y
d
a
tas
e
t.
Ba
se
d
o
n
th
e
DBL
P
ra
w
d
a
ta,
th
e
c
las
sifica
ti
o
n
p
ro
c
e
ss
is
p
r
o
p
o
se
d
b
y
u
sin
g
d
e
e
p
n
e
u
ra
l
n
e
two
rk
s
(DN
Ns
).
In
t
h
e
c
las
sifica
ti
o
n
p
ro
c
e
ss
,
t
h
e
D
BLP
ra
w
d
a
ta
d
iv
id
e
d
in
t
o
f
iv
e
fe
a
tu
re
s,
in
c
lu
d
in
g
n
a
m
e
,
a
u
th
o
r,
ti
tl
e
,
v
e
n
u
e
,
a
n
d
y
e
a
r.
Twe
lv
e
sc
e
n
a
rio
s
a
re
d
e
sig
n
e
d
with
a
d
iffere
n
t
stru
c
t
u
re
to
v
a
li
d
a
te
a
n
d
se
lec
t
th
e
b
e
st
m
o
d
e
l
o
f
DN
Ns
.
F
u
rth
e
rm
o
re
,
t
h
is
p
a
p
e
r
is
a
lso
c
o
m
p
a
re
d
DN
Ns
with
o
th
e
r
c
las
si
fiers
,
su
c
h
a
s
su
p
p
o
rt
v
e
c
to
r
m
a
c
h
i
n
e
(S
VM)
a
n
d
d
e
c
isio
n
tree
.
Th
e
re
su
lt
s
sh
o
w
DN
Ns
o
u
t
p
e
rfo
rm
S
VM
a
n
d
d
e
c
isio
n
tr
e
e
m
e
th
o
d
s
in
a
ll
p
e
rfo
rm
a
n
c
e
m
e
tri
c
s.
Th
e
DN
Ns
p
e
rfo
rm
a
n
c
e
s
with
t
h
re
e
h
i
d
d
e
n
lay
e
rs
a
s
t
h
e
b
e
st
m
o
d
e
l,
a
c
h
iev
e
a
c
c
u
ra
c
y
,
se
n
siti
v
it
y
,
sp
e
c
i
ficity
,
p
re
c
isio
n
,
a
n
d
F
1
-
sc
o
re
a
re
9
8
.
8
5
%
,
9
5
.
9
5
%
,
9
9
.
2
6
%
,
9
4
.
8
0
%
,
a
n
d
9
5
.
3
6
%
,
r
e
sp
e
c
ti
v
e
ly
.
In
th
e
fu
t
u
re
,
DN
Ns
a
re
m
o
re
p
e
rfo
rm
in
g
with
th
e
a
u
to
m
a
ted
fe
a
tu
re
re
p
re
se
n
tatio
n
in
AN
D p
ro
c
e
ss
in
g
.
K
ey
w
o
r
d
s
:
Au
th
o
r
n
a
m
e
d
is
am
b
ig
u
atio
n
B
ib
lio
g
r
ap
h
ic
d
ata
Deep
n
eu
r
al
n
etwo
r
k
s
Ho
m
o
n
y
m
Sy
m
o
n
y
m
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r
th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Sit
i N
u
r
m
ain
i
I
n
tellig
en
t Sy
s
tem
s
R
esear
ch
Gr
o
u
p
Un
iv
er
s
itas
Sriwijay
a
Palem
b
an
g
3
0
1
3
7
,
I
n
d
o
n
esia
E
m
ail: sit
i_
n
u
r
m
ain
i@
u
n
s
r
i.c
.
id
1.
I
NT
RO
D
UCT
I
O
N
Au
th
o
r
n
am
e
d
is
am
b
ig
u
atio
n
(
AND)
in
th
e
p
u
b
licatio
n
is
a
well
-
k
n
o
wn
p
r
o
b
lem
.
Su
ch
a
tech
n
iq
u
e
u
s
ed
to
o
v
e
r
co
m
e
th
e
p
r
o
b
le
m
o
f
am
b
ig
u
o
u
s
in
d
ig
ital
lib
r
ar
ies
(
DL
)
,
in
clu
d
in
g
DB
L
P,
Go
o
g
le
Sch
o
lar
,
an
d
o
th
er
s
.
W
h
en
s
ea
r
ch
in
g
f
o
r
a
n
au
th
o
r
’
s
n
a
m
e
o
r
a
r
ticle
ti
tle
with
a
s
p
ec
if
ic
au
th
o
r
’
s
n
am
e
o
n
a
DL
,
am
b
ig
u
ity
p
r
o
b
lem
s
o
f
ten
ar
is
e.
T
h
u
s
m
an
y
r
elate
d
a
r
ticles
will
ap
p
ea
r
with
th
e
s
am
e
n
am
e
o
r
th
e
s
am
e
titl
e
in
a
b
ib
lio
g
r
ap
h
ic
d
atab
ase
[
1
]
.
B
asically
,
AND
co
n
d
itio
n
o
cc
u
r
s
b
ec
au
s
e
o
f
th
e
f
o
llo
wi
n
g
f
o
u
r
r
ea
s
o
n
s
as
f
o
llo
w
[
2
]
;
(
i)
ca
u
s
e
d
b
y
s
o
m
e
o
n
e
wh
o
p
u
b
lis
h
es
with
d
if
f
er
en
t
n
am
es;
(
ii)
m
an
y
au
th
o
r
s
p
u
b
lis
h
with
th
e
s
am
e
n
am
e;
(
iii)
in
co
m
p
lete
d
ata
o
r
er
r
o
r
s
;
an
d
(
iv
)
th
e
in
cr
ea
s
in
g
n
u
m
b
er
o
f
ar
ticles
an
d
o
r
jo
u
r
n
als
p
u
b
lis
h
ed
o
n
DL
.
T
h
ese
f
o
u
r
ca
u
s
es
ca
n
b
e
u
s
ed
as
a
r
ef
er
en
ce
f
o
r
g
at
h
er
i
n
g
all
th
e
in
f
o
r
m
atio
n
n
ee
d
ed
t
o
f
in
d
o
u
t
an
d
s
o
lv
e
th
e
p
r
o
b
lem
o
f
am
b
ig
u
ity
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
N
eu
r
a
l n
etw
o
r
k
tech
n
iq
u
e
w
ith
d
ee
p
s
tr
u
ctu
r
e
fo
r
imp
r
o
vin
g
a
u
th
o
r
…
(
F
ir
d
a
u
s
)
1209
T
o
o
v
e
r
co
m
e
th
e
ab
o
v
e
p
r
o
b
le
m
s
,
s
ev
er
al
r
esear
ch
er
s
h
av
e
p
r
o
p
o
s
ed
a
s
o
l
u
tio
n
[
3
]
–
[
5
]
.
H
o
wev
er
,
th
e
m
eth
o
d
ca
n
’
t
g
u
ar
a
n
tee
ac
cu
r
ate
r
esu
lts
.
T
h
e
AND
to
p
ic
was
in
tr
o
d
u
ce
d
b
y
V.
I
.
T
o
r
v
ik
[
6
]
.
B
u
t
s
p
ec
if
ically
,
th
e
s
u
b
ject
o
f
AND
is
p
u
r
s
ed
in
th
e
ca
s
e
o
f
h
o
m
o
n
y
m
s
an
d
s
y
n
o
n
y
m
s
.
T
h
e
ca
s
e
o
f
h
o
m
o
n
y
m
s
an
d
s
y
n
o
n
y
m
s
is
th
e
co
r
e
p
r
o
b
lem
in
AND
wh
ich
m
a
k
es
it
ev
e
n
m
o
r
e
co
m
p
licated
[
7
]
.
Ho
m
o
n
y
m
s
ar
e
ca
s
es
wh
er
e
two
n
am
es
ar
e
th
e
s
am
e
in
a
j
o
u
r
n
al
p
u
b
licatio
n
b
u
t
a
r
e
o
wn
e
d
b
y
d
if
f
e
r
en
t
p
eo
p
le,
wh
ile
Sy
n
o
n
y
m
s
i
s
th
e
o
p
p
o
s
ite
ca
s
e
wh
en
th
er
e
ar
e
d
if
f
er
e
n
t
n
am
es
b
u
t
ar
e
o
wn
ed
b
y
th
e
s
am
e
p
er
s
o
n
[
8
]
,
[
9
]
.
T
h
e
r
ese
ar
ch
f
o
cu
s
ed
o
n
th
e
h
o
m
o
n
y
m
ca
s
e
was c
o
n
d
u
cted
b
y
Mo
m
en
i
F.
u
s
in
g
DB
L
P b
ib
lio
g
r
ap
h
ic
d
ata
[
1
0
]
.
T
h
e
s
tu
d
y
aim
s
to
ev
alu
ate
th
e
m
eth
o
d
u
s
ed
f
o
r
th
e
n
etwo
r
k
c
o
-
au
t
h
o
r
s
in
th
e
ca
s
e
o
f
h
o
m
o
n
y
m
au
th
o
r
s
b
y
clu
s
ter
i
n
g
o
n
t
h
e
s
am
e
n
am
e
d
ata
(
h
o
m
o
n
y
m
)
.
T
h
e
r
esear
c
h
y
ield
e
d
g
o
o
d
p
er
f
o
r
m
an
ce
f
o
r
t
h
e
m
o
s
t
n
am
es.
Un
f
o
r
tu
n
ately
,
th
e
m
eth
o
d
u
s
ed
s
till
n
ee
d
ed
o
p
tim
izati
o
n
.
Ma
n
y
s
tu
d
ies
o
n
AND
to
p
ics
p
er
tain
to
th
e
h
o
m
o
n
y
m
an
d
s
y
n
o
n
y
m
ca
s
es
[
9
]
,
[
1
1
]
-
[
1
3
]
.
Ho
we
v
er
,
n
o
s
p
ec
if
ic
r
esear
ch
f
o
cu
s
es
o
n
f
in
d
in
g
s
o
lu
tio
n
s
in
th
e
ca
s
es
o
f
h
o
m
o
n
y
m
s
an
d
s
y
n
o
n
y
m
s
class
if
icatio
n
.
C
u
r
r
en
tly
,
th
e
r
e
ar
e
s
ev
e
r
al
m
eth
o
d
s
h
a
v
e
p
r
o
p
o
s
ed
to
g
iv
e
a
s
o
lu
tio
n
in
th
e
AND
class
if
icatio
n
p
r
o
b
lem
b
ased
o
n
m
ac
h
in
e
lea
r
n
in
g
with
th
e
s
u
p
er
v
is
ed
[
1
4
]
an
d
u
n
s
u
p
e
r
v
is
ed
ap
p
r
o
ac
h
[
1
5
]
.
On
e
tech
n
iq
u
e
co
m
m
o
n
l
y
u
s
ed
f
o
r
th
e
class
if
icatio
n
,
an
d
it
h
as
b
ee
n
p
r
o
v
en
to
p
r
o
v
id
e
g
o
o
d
r
esu
lts
is
s
u
p
p
o
r
t
v
ec
to
r
m
ac
h
in
e
(
SVM)
.
J
aso
n
D.
M.
R
en
n
ie
c
o
n
d
u
cte
d
a
m
u
lticlas
s
tex
t
class
if
icatio
n
u
s
in
g
SVM,
an
d
it
c
o
m
p
ar
ed
with
Naïv
e
B
ay
es
[
1
6
]
.
T
h
e
r
esu
lts
p
r
o
v
e
th
at
SVM
ca
n
r
ed
u
ce
lo
s
s
e
s
1
0
%
to
2
0
%
lo
wer
to
Naïv
e
B
ay
es,
wh
ich
m
ea
n
s
SVM
h
as
th
e
p
er
f
o
r
m
an
ce
t
o
r
ed
u
ce
lo
s
s
es
to
th
e
lo
west
p
o
in
t.
J
aso
n
D.
M.
R
en
n
i
e
p
r
esen
t
m
u
lticlas
s
class
if
icatio
n
with
SVM,
a
s
i
m
ilar
s
tu
d
y
was
co
n
d
u
cted
b
y
Giles
M.
Fo
o
d
y
with
im
a
g
e
d
atasets
[
1
7
]
.
I
n
s
u
c
h
a
s
tu
d
y
SVM
co
m
p
ar
es
with
d
is
cr
im
in
an
t
an
aly
s
is
an
d
d
e
cisi
o
n
tr
ee
s
.
T
h
e
r
esu
lts
as
s
h
o
wn
b
y
u
s
in
g
SVM
h
as
9
3
.
7
% a
cc
u
r
ac
y
,
d
is
cr
im
in
an
t
an
aly
s
is
h
as 9
0
% a
cc
u
r
ac
y
,
a
n
d
d
ec
is
io
n
tr
ee
h
as 9
0
.
3
% a
cc
u
r
ac
y
r
esp
ec
tiv
ely
.
On
e
m
eth
o
d
o
f
ar
tific
ial
n
eu
r
a
l
n
etwo
r
k
s
with
a
“d
ee
p
”
s
tr
u
c
tu
r
e
k
n
o
w
n
as
th
e
d
ee
p
n
e
u
r
al
n
etwo
r
k
s
(
DNNs)
m
eth
o
d
is
th
e
m
o
s
t
p
o
p
u
lar
a
n
d
wid
el
y
u
s
ed
i
n
class
if
icatio
n
p
r
o
b
lem
s
.
DNNs
cla
s
s
if
ier
p
r
o
d
u
ce
s
an
ex
ce
llen
t
p
er
f
o
r
m
an
ce
f
o
r
tex
t
p
r
o
ce
s
s
in
g
[
1
8
]
.
I
n
[
1
9
]
,
th
e
DNNs
m
eth
o
d
s
ig
n
if
ican
tly
o
u
tp
er
f
o
r
m
s
o
t
h
er
m
eth
o
d
s
an
d
p
r
o
d
u
ce
s
9
9
.
3
1
%
ac
cu
r
ac
y
in
th
e
Vietn
am
ese
a
u
th
o
r
n
a
m
e.
Ho
wev
er
,
th
is
m
e
th
o
d
o
n
l
y
d
etec
tio
n
au
th
o
r
am
b
ig
u
atio
n
s
,
wh
e
r
ea
s
,
h
o
m
o
n
y
m
s
an
d
s
y
n
o
n
y
m
s
ar
e
an
ess
en
tial p
r
o
b
lem
in
AND.
T
o
th
e
b
est o
f
o
u
r
k
n
o
wled
g
e,
o
n
ly
lim
ited
r
esear
ch
ab
o
u
t
AND
to
p
ic
b
ased
o
n
DNNs
tech
n
iq
u
e
in
th
e
l
iter
atu
r
e,
an
d
s
u
ch
r
esear
ch
with
o
u
t
in
v
esti
g
atio
n
in
h
o
m
o
n
y
m
a
n
d
s
y
n
o
n
y
m
ca
s
e.
Hen
ce
,
th
e
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
class
if
icatio
n
is
d
esira
b
le
to
th
e
d
ee
p
in
v
esti
g
atio
n
.
T
h
e
r
est
o
f
th
e
p
ap
er
i
s
s
tr
u
ctu
r
ed
as
f
o
llo
ws.
I
n
th
e
in
tr
o
d
u
ctio
n
,
s
o
m
e
r
elate
d
wo
r
k
s
to
AND
ar
e
d
is
cu
s
s
ed
,
an
d
th
eir
ca
p
ab
ilit
ie
s
an
d
lim
itatio
n
s
ar
e
h
ig
h
lig
h
ted
.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
f
DNNs
d
escr
ib
es
i
n
d
etail
t
h
e
wo
r
k
in
g
o
f
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
class
if
icatio
n
.
W
e
s
im
u
lated
th
e
p
r
o
p
o
s
ed
alg
o
r
it
h
m
o
n
th
e
J
i
n
s
eo
k
Kim
d
ataset
an
d
co
m
p
ar
e
it
to
b
aselin
e
m
eth
o
d
s
i
n
ex
p
er
i
m
en
ts
an
d
d
is
cu
s
s
io
n
s
.
I
n
th
e
en
d
,
we
co
n
clu
d
e
with
a
d
is
cu
s
s
io
n
o
f
co
n
clu
s
io
n
s
an
d
f
u
tu
r
e
wo
r
k
in
c
o
n
clu
s
io
n
s
.
2.
M
AT
E
R
I
AL
A
ND
M
E
T
H
O
D
S
T
h
is
p
ap
er
p
r
o
p
o
s
es
th
e
tex
t
p
r
o
ce
s
s
in
g
m
eth
o
d
to
ca
lcu
late
ap
p
r
o
p
r
iate
f
ea
tu
r
es
f
r
o
m
th
e
AND
r
aw
d
ata.
T
h
e
m
eth
o
d
co
n
s
is
ts
o
f
d
ata
ac
q
u
is
i
tio
n
,
d
ata
p
r
ep
ar
atio
n
,
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
,
f
ea
tu
r
e
r
ed
u
ctio
n
,
class
if
icatio
n
,
an
d
v
alid
atio
n
,
as in
Fig
u
r
e
1
.
All th
e
s
tag
es c
an
b
e
d
escr
ib
e
d
as d
etail
in
th
e
f
o
llo
win
g
s
ec
tio
n
.
Fig
u
r
e
1
.
AND
class
if
icatio
n
p
r
o
ce
s
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
4
,
Au
g
u
s
t 2
0
2
1
:
1
2
0
8
-
1
2
1
7
1210
2
.
1
.
Da
t
a
a
cquis
it
io
n
T
h
e
d
ataset
u
s
ed
in
th
is
s
tu
d
y
ca
m
e
f
r
o
m
J
in
s
eo
k
Kim
[
2
]
,
with
r
aw
d
ata
m
a
d
e
b
y
Giles
L
ee
[
2
0
]
.
I
n
itially
,
th
e
r
aw
d
ata
co
n
tain
ed
s
ep
ar
ate
n
am
e
d
ata
f
o
r
o
n
e
au
th
o
r
in
o
n
e
f
ile
f
r
o
m
m
a
n
y
Dig
ital
L
ib
r
ar
ies
(
DL
)
.
T
h
e
d
ataset
h
as
u
s
ed
b
y
R
ich
ar
d
o
G.
C
o
ta
[
2
1
]
an
d
Ala
n
F.
San
tan
a
[
2
2
]
.
T
h
e
n
,
th
e
r
a
w
d
ata
is
im
p
r
o
v
ed
b
y
Mu
ller
[
2
3
]
an
d
h
e
f
o
u
n
d
th
at
th
e
d
ataset
h
as
a
ten
d
en
cy
to
war
d
s
DL
DB
L
P
an
d
m
atc
h
es
it
with
D
B
L
P
d
ata.
Fu
r
th
er
m
o
r
e
,
th
e
d
ataset
was
clea
n
ed
u
p
ag
ain
an
d
co
m
p
let
ed
b
y
J
in
s
eo
k
Kim
[
8
]
b
y
a
d
d
i
n
g
m
is
s
in
g
d
ata
a
n
d
elim
in
atin
g
u
n
n
ec
ess
ar
y
d
ata
b
ased
o
n
DB
L
P
d
ata
s
o
th
at
th
e
d
ataset
b
ec
o
m
es
a
DB
L
P
d
ataset
with
clea
r
en
tit
ies
an
d
attr
ib
u
tes.
I
n
th
is
d
ataset,
th
er
e
ar
e
s
ix
attr
ib
u
te
s
alwa
y
s
u
s
ed
b
ec
au
s
e
it
h
as
a
g
r
ea
t
in
f
lu
en
ce
o
n
AND
d
ata.
T
h
e
s
ix
attr
ib
u
tes
ar
e
a
n
am
e,
au
th
o
r
,
au
t
h
o
r
I
D,
titl
e,
v
en
u
e,
a
n
d
y
ea
r
.
T
h
is
d
ataset
h
as
th
e
s
am
e
len
g
th
o
r
a
m
o
u
n
t
o
f
d
ata
f
o
r
ea
ch
at
tr
ib
u
te
ab
o
u
t
5
0
1
8
d
ata.
T
h
er
e
ar
e
4
5
6
d
if
f
er
en
t
n
a
m
es
in
th
e
n
am
e
attr
ib
u
te
with
th
r
ee
n
am
es
o
v
er
1
0
0
.
T
h
e
m
o
s
t
n
am
es
ar
e
Seo
n
g
-
W
h
an
L
ee
ab
o
u
t
1
2
5
,
Dav
id
S.
J
o
h
n
s
o
n
ab
o
u
t
1
0
4
,
an
d
An
o
o
p
Gu
p
ta
ab
o
u
t
1
0
4
.
W
h
ile
in
th
e
Au
th
o
r
s
attr
ib
u
te
th
e
r
e
ar
e
as
m
an
y
as
4
6
5
4
n
a
m
e
s
d
if
f
er
en
t,
wh
ich
s
h
o
ws
th
at
th
e
n
am
e
in
t
h
e
A
u
th
o
r
s
attr
ib
u
te
is
v
e
r
y
m
u
ch
d
if
f
er
en
t
in
n
u
m
b
er
f
r
o
m
t
h
e
n
am
e
in
th
e
Nam
e
attr
ib
u
te.
T
h
is
is
d
u
e
to
th
e
f
a
ct
th
at
in
a
jo
u
r
n
al
p
u
b
licatio
n
,
th
er
e
ar
e
m
a
n
y
au
t
h
o
r
s
wh
o
f
o
llo
w
th
e
au
th
o
r
’
s
n
am
e
in
th
e
Nam
e
attr
ib
u
te.
W
h
er
ea
s
th
e
Au
th
o
r
I
D
attr
ib
u
te
h
as
4
8
0
d
is
tin
ct
d
ata
o
f
Au
th
o
r
I
D
d
ata.
I
t
m
ea
n
s
th
er
e
ar
e
4
8
0
d
if
f
er
en
t la
b
els s
ca
tter
ed
in
th
e
d
ata
alo
n
g
5
0
1
8
.
2
.
2
.
Da
t
a
prepa
ra
t
i
o
n
All
attr
ib
u
tes
co
n
tain
ed
in
th
e
d
ataset
will
b
e
s
ep
ar
ated
in
to
f
ea
tu
r
e
attr
ib
u
tes
an
d
lab
el
attr
ib
u
tes.
T
h
e
f
ea
tu
r
e
attr
ib
u
te
is
co
m
m
o
n
ly
r
ef
er
r
ed
to
as
in
p
u
t,
wh
ich
is
an
attr
ib
u
te
th
at
will
b
e
u
s
ed
as
in
p
u
t
d
ata
to
b
e
p
r
o
ce
s
s
ed
b
y
th
e
class
if
ier
.
T
h
e
r
esu
lts
will
b
e
g
r
o
u
p
ed
in
t
o
o
n
e
o
f
t
h
e
d
ata
in
th
e
la
b
el
o
r
o
u
tp
u
t
att
r
ib
u
te.
T
h
e
lab
el
attr
ib
u
te
is
an
attr
ib
u
te
th
at
is
s
elec
ted
f
r
o
m
m
an
y
att
r
ib
u
tes
th
at
ex
is
t
in
th
e
d
ata
th
at
will
b
e
u
s
ed
as
o
u
tp
u
t
(
a
p
lace
to
g
r
o
u
p
in
p
u
t
d
ata)
f
r
o
m
f
ea
tu
r
e
o
r
in
p
u
t
at
tr
ib
u
tes.
Featu
r
e
attr
ib
u
tes
ar
e
tak
en
f
r
o
m
all
th
e
attr
ib
u
tes
th
at
h
a
v
e
im
p
o
r
tan
t
in
f
lu
en
ce
s
in
th
e
d
ata
asid
e
f
r
o
m
th
e
la
b
el
attr
ib
u
tes.
T
h
e
lab
el
attr
ib
u
te
is
tak
e
n
f
r
o
m
th
e
m
o
s
t
s
p
ec
if
ic
an
d
th
e
u
n
iq
u
e
attr
ib
u
te
am
o
n
g
all
th
e
attr
ib
u
tes
in
th
e
d
ata
th
at
will
b
e
ab
le
to
d
is
tin
g
u
is
h
th
e
d
ata
g
r
o
u
p
s
th
at
ar
e
class
if
ied
.
I
n
th
is
s
tu
d
y
,
th
e
Au
th
o
r
I
D
attr
ib
u
te
is
u
s
ed
as
a
lab
el
attr
ib
u
te
b
ec
au
s
e
it
is
s
p
ec
if
ic
an
d
u
n
iq
u
e.
I
n
d
ata
p
r
e
p
ar
atio
n
,
t
h
e
m
ain
t
ask
is
to
f
in
d
in
d
iv
id
u
al
h
o
m
o
g
en
eo
u
s
d
ata
in
th
e
d
ataset
an
d
s
to
r
e
it
as
a
n
ew
lab
el
(
s
ee
in
Fig
u
r
e
2
)
.
B
y
ad
d
in
g
a
h
o
m
o
n
y
m
lab
el
c
o
lu
m
n
t
o
th
e
d
ataset
th
at
ca
n
b
e
in
itialized
lab
el
1
f
o
r
t
h
e
h
o
m
o
n
y
m
an
d
lab
el
0
f
o
r
t
h
e
n
o
n
-
h
o
m
o
n
y
m
.
T
h
e
n
e
x
t
s
tep
is
to
f
in
d
th
e
s
y
n
o
n
y
m
lab
els
in
t
h
e
s
am
e
way
.
T
h
e
r
esu
lt
is
ad
d
ed
as
a
n
ew
lab
el
co
lu
m
n
with
in
itia
lizatio
n
lab
el
1
f
o
r
s
y
n
o
n
y
m
d
at
a
an
d
lab
el
0
f
o
r
non
-
s
y
n
o
n
y
m
.
Fu
r
t
h
er
m
o
r
e,
t
wo
co
lu
m
n
s
a
r
e
ad
d
ed
in
th
e
d
ataset,
th
e
f
ir
s
t c
o
lu
m
n
o
f
th
e
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
with
lab
el
1
lab
els
a
n
d
n
o
n
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
with
lab
el
0
.
T
h
e
s
ec
o
n
d
is
a
n
o
n
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
lab
el
co
lu
m
n
,
wh
ich
is
a
lab
el
d
er
iv
ed
f
r
o
m
d
ata
th
at
is
n
o
t
in
clu
d
ed
in
t
h
e
h
o
m
o
n
y
m
la
b
el
co
lu
m
n
an
d
a
s
y
n
o
n
y
m
o
r
th
e
o
p
p
o
s
ite
o
f
t
h
e
s
y
n
o
n
y
m
lab
el
co
lu
m
n
.
T
h
er
e
f
o
r
e,
f
o
u
r
n
ew
lab
el
co
l
u
m
n
s
ar
e
ac
h
iev
ed
,
ea
ch
n
ew
lab
el
co
lu
m
n
will
b
e
m
er
g
ed
in
to
a
n
ew
lab
el
c
o
lu
m
n
with
lab
els 0
,
1
,
2
,
an
d
3
r
e
p
r
esen
tin
g
ea
c
h
n
ew
lab
el
c
o
lu
m
n
h
o
m
o
n
y
m
,
s
y
n
o
n
y
m
,
h
o
m
o
n
y
m
,
n
o
n
-
s
y
n
o
n
y
m
r
esp
ec
tiv
ely
.
T
h
u
s
,
th
e
tr
ain
in
g
to
b
e
p
er
f
o
r
m
ed
is
a
f
o
u
r
-
lab
el
tr
ain
in
g
th
at
r
esu
lts
f
r
o
m
th
e
s
ea
r
ch
f
o
r
h
o
m
o
n
y
m
s
,
s
y
n
o
n
y
m
s
,
s
y
n
o
n
y
m
s
,
a
n
d
n
o
n
-
s
y
n
o
n
y
m
s
.
Fig
u
r
e
2
.
Data
p
r
ep
ar
atio
n
p
r
o
ce
s
s
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
N
eu
r
a
l n
etw
o
r
k
tech
n
iq
u
e
w
ith
d
ee
p
s
tr
u
ctu
r
e
fo
r
imp
r
o
vin
g
a
u
th
o
r
…
(
F
ir
d
a
u
s
)
1211
2
.
3
.
Da
t
a
pre
-
pro
ce
s
s
ing
I
n
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
s
tag
e,
al
l
f
ea
tu
r
es
ex
ce
p
t
th
e
Yea
r
f
ea
t
u
r
e
ar
e
ch
an
g
ed
to
d
u
m
m
y
v
ar
iab
le
.
T
h
e
p
r
o
ce
s
s
o
f
r
aw
d
ata
p
r
o
ce
s
s
in
g
in
to
a
d
u
m
m
y
v
ar
iab
le
ca
n
b
e
r
ep
r
esen
ted
in
Fig
u
r
e
3
.
T
h
ese
f
ea
tu
r
es
ca
n
b
e
ea
s
ily
in
ter
p
r
eted
b
y
th
e
class
if
ier
.
Fo
r
all
f
ea
tu
r
es
in
th
e
f
o
r
m
o
f
tex
t,
it
will
b
e
ch
an
g
e
d
in
to
n
u
m
er
ic
f
o
r
m
.
Fo
r
ex
am
p
le,
t
h
e
Nam
e
f
ea
tu
r
e
is
ch
an
g
ed
f
r
o
m
n
am
e
te
x
t
to
a
n
u
m
b
e
r
with
a
s
p
ec
if
ic
v
alu
e
o
r
n
u
m
b
er
in
ea
c
h
n
am
e,
an
d
th
e
v
alu
e
w
ill
b
e
th
e
s
am
e
f
o
r
th
e
s
am
e
n
am
e.
T
h
e
v
alu
e
is
g
iv
en
in
a
r
an
d
o
m
v
alu
e.
I
f
th
er
e
is
a
n
am
e
lik
e
“G
u
p
ta”,
t
h
en
t
h
e
n
a
m
e
will
b
e
r
e
p
r
esen
ted
as
a
r
a
n
d
o
m
n
u
m
b
e
r
f
r
o
m
0
t
o
9
.
T
h
at
n
u
m
b
er
will
d
if
f
e
r
as
m
u
ch
as
th
e
d
if
f
er
en
t
n
am
es
th
at
ex
is
t.
I
f
th
er
e
ar
e
4
5
6
d
is
tin
ct
n
am
es,
th
en
th
er
e
w
ill
b
e
4
5
6
d
if
f
er
en
t
n
u
m
b
er
s
,
w
h
ich
will
r
ep
r
esen
t
ea
ch
n
am
e
in
o
n
e
co
lu
m
n
alo
n
g
5
0
1
8
d
ata.
Af
ter
all,
n
am
e
s
b
ec
o
m
e
n
u
m
b
er
s
an
d
s
to
r
ed
in
a
v
ar
iab
le,
a
n
ew
v
ar
iab
le
will
b
e
p
r
ep
ar
ed
to
s
to
r
e
d
u
m
m
y
v
ar
iab
les
wh
o
s
e
co
n
ten
ts
ar
e
0
an
d
1
.
T
h
e
n
ew
v
ar
iab
le
will
co
n
tain
a
n
u
m
b
er
o
f
co
lu
m
n
s
th
e
s
am
e
n
u
m
b
er
o
f
th
e
n
am
es.
I
f
th
er
e
ar
e
4
5
6
d
if
f
er
en
t
n
am
es,
th
en
th
er
e
will
b
e
4
5
6
co
lu
m
n
s
in
th
e
n
ew
v
ar
iab
le.
T
h
e
len
g
th
o
f
th
e
d
ata
o
n
th
e
n
ew
v
ar
iab
le
will
b
e
th
e
s
am
e
as th
e
p
r
ev
io
u
s
v
ar
ia
b
l
e
ab
o
u
t
5
0
1
8
d
ata
r
o
ws co
n
t
ain
in
g
0
a
n
d
1
.
Fig
u
r
e
3
.
Data
p
r
e
-
p
r
o
ce
s
s
in
g
T
h
e
n
u
m
b
er
r
ep
r
esen
tin
g
ea
ch
n
am
e
in
th
e
f
ir
s
t
v
a
r
iab
le
will
b
e
u
s
ed
as
t
h
e
in
d
e
x
co
l
u
m
n
i
n
th
e
n
e
w
v
ar
iab
le.
I
f
t
h
e
n
am
e
“Gu
p
ta”
is
r
ep
r
esen
ted
b
y
th
e
n
u
m
b
er
9
,
th
en
in
th
e
n
in
th
co
lu
m
n
in
th
e
n
ew
v
ar
iab
le
h
as
th
e
v
alu
e
1
an
d
th
e
o
th
er
co
lu
m
n
will
b
e
0
.
T
h
r
o
u
g
h
o
u
t
th
e
n
in
th
co
l
u
m
n
9
f
r
o
m
t
h
e
f
ir
s
t
r
o
w
to
th
e
5
0
1
8
th
r
o
w
g
iv
en
a
v
alu
e
o
f
1
,
if
th
e
p
r
e
v
io
u
s
v
ar
iab
le
co
n
tain
s
th
e
n
u
m
b
er
9
.
Fo
r
ex
am
p
le,
if
i
n
th
e
f
ir
s
t
v
ar
iab
le
th
e
n
u
m
b
er
9
in
t
h
e
in
d
e
x
r
o
w
f
i
f
th
an
d
h
u
n
d
r
ed
th
,
th
en
in
th
e
n
in
th
co
lu
m
n
,
th
e
f
if
t
h
an
d
h
u
n
d
r
ed
th
r
o
ws
in
th
e
n
ew
v
ar
iab
le
will
b
e
wo
r
th
1
a
n
d
o
t
h
er
wis
e
0
.
T
h
e
r
u
le
will a
p
p
ly
to
all
d
if
f
e
r
en
t
n
u
m
b
e
r
s
in
th
e
n
ew
v
ar
ia
b
le.
Pre
-
p
r
o
ce
s
s
in
g
th
e
au
th
o
r
’
s
f
ea
tu
r
e
is
d
if
f
er
en
t
f
r
o
m
th
e
Nam
e
f
ea
tu
r
e,
b
u
t
it
is
s
till
im
p
o
r
tan
t
to
cr
ea
te
a
d
u
m
m
y
v
ar
iab
le.
T
h
e
s
tep
is
to
s
ep
ar
ate
ea
ch
tex
t
b
y
n
am
e
f
o
r
all
th
e
n
am
es
in
th
e
Au
th
o
r
s
f
ea
t
u
r
e
an
d
s
av
e
it
to
a
v
ar
iab
le
(
e.
g
.
v
a
r
iab
le
“x
”)
.
T
h
en
a
d
ictio
n
ar
y
is
also
m
ad
e
wh
ich
is
s
to
r
ed
in
a
v
ar
iab
le
(
e.
g
.
v
ar
iab
le
“z
”)
f
r
o
m
th
e
n
am
e
d
ata
in
v
ar
iab
le
x
.
I
n
th
e
d
esig
n
o
f
a
d
ictio
n
a
r
y
d
o
es
n
o
t
ch
an
g
e
th
e
s
h
ap
e
o
r
n
u
m
b
er
o
f
d
ata
in
t
h
e
“x
”
v
ar
iab
le.
T
h
e
n
all
th
e
n
am
es
in
th
e
v
ar
iab
le
“x
”
ar
e
g
iv
en
th
e
v
alu
e
0
,
an
d
also
s
to
r
ed
in
a
v
ar
iab
le
(
e.
g
.
v
ar
iab
le
“t”)
.
T
h
e
“x
”
v
ar
ia
b
le
is
co
m
p
ar
ed
to
th
e
“z
”
v
ar
iab
le
b
y
r
ef
er
r
in
g
to
th
e
“t”
v
ar
iab
le,
an
d
th
e
o
u
tp
u
t
o
f
th
e
p
r
o
ce
s
s
is
a
d
u
m
m
y
v
a
r
iab
le
as
a
f
ea
tu
r
e
to
b
e
an
in
p
u
t
class
if
ier
.
B
asically
,
th
e
p
r
o
ce
s
s
ch
an
g
es
th
e
v
alu
e
o
f
0
in
th
e
v
ar
iab
le
“t”
t
o
1
,
if
th
e
v
ar
iab
le
“z
”
f
i
n
d
s
t
h
e
s
am
e
n
am
e
i
n
th
e
v
ar
ia
b
le
“x
”
.
Pre
-
p
r
o
ce
s
s
in
g
o
f
th
e
Au
t
h
o
r
s
,
T
itle
an
d
v
e
n
u
e
f
ea
tu
r
es
ar
e
d
o
n
e
in
th
e
s
am
e
way
b
ec
a
u
s
e
it
h
as
th
e
s
am
e
f
o
r
m
o
f
d
ata.
Ho
wev
er
,
p
r
e
-
p
r
o
ce
s
s
in
g
in
th
e
y
ea
r
f
ea
tu
r
e
h
as
a
d
if
f
er
en
t
s
tep
f
r
o
m
t
h
e
p
r
ev
io
u
s
f
o
u
r
f
e
atu
r
es,
b
ec
au
s
e
th
e
Yea
r
f
ea
tu
r
e
is
n
u
m
er
ic,
it
o
n
l
y
n
ee
d
s
to
b
e
n
o
r
m
alize
d
in
to
a
v
alu
e
o
n
a
s
ca
le
o
f
0
to
1
.
Fin
ally
,
all
f
ea
tu
r
es
ar
e
co
llected
in
to
a
s
in
g
le
d
ata
w
ith
lab
els
as
th
e
in
itial
d
ataset.
T
h
en
,
th
e
f
ea
tu
r
e
d
ata
is
d
iv
id
ed
in
to
8
0
%
f
o
r
tr
ain
in
g
an
d
2
0
% f
o
r
test
in
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
4
,
Au
g
u
s
t 2
0
2
1
:
1
2
0
8
-
1
2
1
7
1212
2
.
4
.
Cla
s
s
if
ica
t
io
n
w
it
h dee
p neura
l net
wo
rk
s
I
n
th
is
s
tu
d
y
,
th
e
m
u
lti
-
class
-
class
if
icatio
n
o
f
AND
b
ased
o
n
DNNs
u
s
e
th
e
im
p
lem
en
tatio
n
o
f
th
e
b
ac
k
p
r
o
p
ag
atio
n
(
B
P)
alg
o
r
ith
m
f
o
r
tr
ain
i
n
g
an
d
test
in
g
.
T
h
e
class
if
ier
is
tr
ain
ed
b
y
u
s
in
g
in
p
u
t
d
ata
x
a
n
d
an
an
n
o
tated
la
b
el
as a
n
o
u
tp
u
t.
T
h
e
So
f
tm
ax
f
u
n
ctio
n
is
u
s
ed
as a
n
ac
tiv
atio
n
f
u
n
ctio
n
f
o
r
th
e
o
u
tp
u
t la
y
er
o
f
th
e
class
if
ier
.
B
y
u
s
in
g
th
e
So
f
tm
a
x
f
u
n
ctio
n
,
th
e
o
u
tp
u
t
o
f
ea
c
h
u
n
it
ca
n
b
e
tr
ea
ted
as
th
e
p
r
o
b
a
b
ilit
y
o
f
ea
ch
lab
el.
Her
e,
let
N
b
e
th
e
n
u
m
b
er
o
f
u
n
its
o
f
th
e
o
u
tp
u
t
lay
er
,
let
x
b
e
th
e
in
p
u
t,
an
d
let
b
e
th
e
o
u
tp
u
t
o
f
u
n
it
i.
T
h
en
,
th
e
o
u
tp
u
t
(
)
o
f
u
n
it I
is
d
e
f
in
ed
b
y
in
(
1
)
:
(
)
=
∑
=
1
(
1
)
C
r
o
s
s
en
tr
o
p
y
is
u
s
ed
as th
e
lo
s
s
f
u
n
ctio
n
o
f
t
h
e
class
if
ier
L
F a
s
f
o
llo
w:
(
)
=
−
1
∑
∑
(
)
=
1
=
1
(
2
)
wh
er
e
n
is
th
e
s
am
p
le
s
ize,
m
is
th
e
n
u
m
b
er
o
f
class
es,
is
th
e
o
u
t
p
u
t
o
f
th
e
class
if
ier
o
f
class
j
o
f
th
e
ℎ
s
am
p
le
an
d
is
th
e
an
n
o
tated
la
b
el
o
f
class
j o
f
th
e
ℎ
s
am
p
le.
I
n
th
e
ex
p
er
im
en
t,
all
class
es
a
r
e
d
iv
i
d
ed
in
to
f
o
u
r
ca
teg
o
r
ies,
s
u
ch
as
h
o
m
o
n
y
m
,
s
y
n
o
n
y
m
,
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
,
an
d
n
o
n
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
.
B
ased
o
n
f
o
u
r
class
es,
a
DNNs
co
m
p
r
is
es
m
u
ltip
l
e
n
o
d
es
co
n
n
ec
ted
to
ea
ch
o
th
e
r
,
with
ea
ch
n
o
d
e
r
ep
r
esen
tin
g
th
e
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
e
e
x
p
er
im
e
n
t
is
co
n
d
u
cted
b
y
i
n
cr
ea
s
in
g
th
e
n
u
m
b
er
o
f
h
i
d
d
en
lay
er
s
f
r
o
m
lay
e
r
1
to
lay
er
4
,
an
d
all
th
e
p
e
r
f
o
r
m
an
ce
s
ar
e
o
b
s
er
v
ed
in
o
r
d
e
r
to
ch
o
o
s
e
th
e
b
est
m
o
d
el.
T
h
e
d
ee
p
s
tr
u
ctu
r
es
o
f
NNs
h
av
e
1
0
0
n
o
d
es
in
ea
c
h
lay
e
r
.
T
h
e
ac
tiv
atio
n
f
u
n
ctio
n
u
s
ed
in
th
e
in
p
u
t
lay
er
an
d
at
ea
ch
h
id
d
e
n
lay
er
is
R
eL
U,
wh
ile
th
e
ac
tiv
atio
n
f
u
n
ctio
n
u
s
ed
in
th
e
o
u
t
p
u
t
lay
er
is
So
f
tm
ax
.
I
n
th
e
DNN
s
tr
u
ctu
r
e
lo
s
s
f
u
n
ctio
n
is
ca
teg
o
r
ical
cr
o
s
s
-
en
t
r
o
p
y
with
Ad
am
o
p
tim
izer
.
All
ex
p
er
im
en
ts
ar
e
co
n
d
u
cte
d
with
a
tu
n
i
n
g
lea
r
n
in
g
r
ate
f
r
o
m
0
.
0
0
0
1
d
ec
r
ea
s
e
s
to
0
.
1
,
with
a
b
atch
s
ize
in
c
r
ea
s
e
f
r
o
m
8
to
6
4
with
5
0
ep
o
c
h
s
.
T
h
e
p
a
r
am
ete
r
f
ix
es
f
o
r
ea
c
h
ex
p
e
r
im
en
t
s
tar
tin
g
with
1
h
i
d
d
en
-
la
y
er
with
b
atch
s
ize
8
u
p
t
o
4
h
id
d
en
-
la
y
er
s
an
d
b
atch
s
ize
6
4
.
T
h
e
p
r
o
p
o
s
ed
DNNs stru
ctu
r
e
ca
n
b
e
s
ee
n
in
Fig
u
r
e
4
.
Fig
u
r
e
4
.
DNNs
s
tr
u
ctu
r
e
2
.
5
.
M
o
del
ev
a
lua
t
i
o
n
I
n
th
e
ev
alu
atio
n
s
tag
e
o
f
th
e
m
o
d
el
th
at
h
as
b
ee
n
b
u
ilt,
it
is
lo
o
k
in
g
f
o
r
th
e
v
alu
e
o
f
p
r
e
d
ictin
g
test
in
g
f
ea
tu
r
es
an
d
test
in
g
lab
els
o
f
th
e
m
o
d
el
b
u
ilt,
th
en
u
tili
zin
g
t
h
e
p
r
ed
icted
v
alu
e
o
b
tain
ed
t
o
o
b
tain
th
e
co
n
f
u
s
io
n
m
atr
ix
o
f
th
e
m
o
d
el.
T
h
e
class
if
ier
m
u
s
t
b
e
p
r
o
d
u
ce
d
th
e
t
r
u
s
t
v
alu
e
o
f
th
e
p
r
ed
ictio
n
r
esu
lts
.
T
h
e
class
if
ier
m
ak
es
co
r
r
ec
t
p
r
e
d
ictio
n
s
ev
e
n
th
o
u
g
h
it
is
r
o
u
g
h
ly
in
ter
p
r
eted
as
a
p
r
o
b
a
b
ilit
y
.
Ho
wev
e
r
,
th
e
p
o
s
s
ib
ilit
y
o
f
g
ettin
g
th
e
co
r
r
e
ct
p
r
e
d
ictio
n
i
s
n
o
t
en
o
u
g
h
to
o
n
l
y
g
i
v
e
o
n
e
n
u
m
b
er
.
T
h
e
f
iv
e
m
e
asu
r
es
o
f
o
n
e
b
ea
t
H
ar
e
as
s
h
o
wn
in
(
3
)
to
(
7
)
as f
o
llo
ws:
(
)
=
∑
=
1
∑
=
1
+
∑
=
1
(
3
)
(
)
=
∑
=
1
∑
=
1
+
∑
=
1
(
4
)
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
N
eu
r
a
l n
etw
o
r
k
tech
n
iq
u
e
w
ith
d
ee
p
s
tr
u
ctu
r
e
fo
r
imp
r
o
vin
g
a
u
th
o
r
…
(
F
ir
d
a
u
s
)
1213
(
)
=
∑
=
1
+
∑
=
1
∑
=
1
+
∑
=
1
+
∑
+
∑
=
1
=
1
(
5
)
K
is
th
e
n
u
m
b
er
o
f
b
ea
t
ty
p
es.
T
P
H
(
tr
u
e
p
o
s
itiv
es)
is
th
e
n
u
m
b
e
r
o
f
H
ty
p
es
th
at
ar
e
co
r
r
ec
tl
y
class
if
ied
.
T
N
H
(
tr
u
e
n
e
g
ativ
e)
is
th
e
n
u
m
b
er
o
f
n
o
t
-
H
ty
p
es
th
at
ar
e
co
r
r
ec
tly
class
if
ied
.
FP
H
(
f
alse
p
o
s
itiv
e)
is
th
e
n
u
m
b
er
o
f
n
o
t
-
H
ty
p
es
th
at
ar
e
in
co
r
r
ec
tly
p
r
e
d
icted
as
H
ty
p
es.
FN
H
(
f
alse
n
eg
ativ
e)
is
th
e
n
u
m
b
er
o
f
H
ty
p
es th
at
ar
e
in
co
r
r
ec
tly
p
r
ed
i
cted
as n
o
t
-
H
ty
p
es.
3.
RE
SU
L
T
S
AND
D
I
SCU
SS
I
O
N
T
h
e
f
o
u
r
class
es
o
f
AND
ar
e
class
if
ied
with
D
NNs.
T
h
e
cl
ass
es
co
n
s
i
s
t
o
f
non
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
(
class
0
)
,
h
o
m
o
n
y
m
(
class
1
)
,
s
y
n
o
n
y
m
(
class
2
)
,
a
n
d
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
(
class
3
)
.
T
h
e
twelv
e
m
o
d
els
ar
e
f
in
e
-
tu
n
e
d
with
th
e
d
if
f
er
en
t
b
atch
s
izes
an
d
h
id
d
en
lay
er
s
i
n
ea
ch
f
o
u
r
class
.
On
e
h
id
d
en
la
y
er
is
im
p
lem
e
n
ted
in
th
e
f
ir
s
t
to
f
o
u
r
th
m
o
d
el.
T
h
e
f
o
u
r
th
t
o
eig
h
t
m
o
d
el
u
s
es
two
h
id
d
en
la
y
er
s
.
L
ast,
t
h
e
n
in
th
to
twelf
th
m
o
d
e
l
u
s
es
th
r
ee
h
id
d
en
lay
e
r
s
.
E
ac
h
h
id
d
en
lay
e
r
co
n
s
is
ts
o
f
a
b
at
ch
s
ize
o
f
8
,
1
6
,
3
2
,
an
d
6
4
,
r
e
s
p
ec
tiv
ely
.
Fo
r
th
e
non
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
r
e
p
r
e
s
en
ted
in
T
ab
le
1
,
o
v
er
all,
th
e
g
o
o
d
p
er
f
o
r
m
a
n
ce
with
th
e
p
er
ce
n
tag
e
u
p
9
5
%
h
av
e
wo
r
k
ed
o
u
t
in
twelv
e
m
o
d
els
o
f
DNNs.
T
h
e
b
est
m
o
d
el
ca
n
b
e
s
ee
n
in
th
e
s
ev
en
t
h
an
d
eig
h
t
h
m
o
d
el,
wh
ich
u
s
e
two
h
i
d
d
e
n
lay
er
s
a
n
d
b
atch
s
izes
3
2
an
d
6
4
,
r
esp
e
ctiv
ely
.
T
h
e
d
if
f
er
en
ce
s
b
etwe
en
th
ese
two
m
o
d
els
ar
e
n
o
t sig
n
if
ican
t.
T
ab
le
1
.
DNNs
p
er
f
o
r
m
an
ce
f
o
r
n
o
n
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
cl
ass
if
icatio
n
(
class
0)
M
o
d
e
l
P
e
r
f
o
r
ma
n
c
e
E
v
a
l
u
a
t
i
o
n
(
%)
N
u
mb
e
r
o
f
D
a
t
a
A
c
c
u
r
a
c
y
S
e
n
s
i
t
i
v
i
t
y
S
p
e
c
i
f
i
c
i
t
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
1
98
97
98
99
98
6
1
5
2
96
95
98
99
97
6
0
0
3
97
96
99
99
98
5
9
0
4
97
95
99
99
97
5
9
9
5
98
97
99
99
98
6
0
0
6
97
97
98
98
96
5
8
8
7
98
97
99
99
98
5
9
9
8
98
97
99
99
98
6
0
6
9
97
96
99
99
97
5
9
0
10
97
96
98
99
97
6
0
9
11
97
96
98
98
97
6
1
1
12
97
96
99
99
98
5
8
4
Fo
r
th
e
h
o
m
o
n
y
m
class
,
T
ab
le
2
s
h
o
ws
th
e
r
esu
lts
o
f
u
p
to
8
7
%
as
o
v
er
all
in
th
e
twelv
e
m
o
d
els.
T
h
e
n
u
m
b
er
o
f
h
o
m
o
n
y
m
d
ata
s
m
aller
th
an
th
e
non
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
class
.
T
h
e
f
if
th
m
o
d
el
with
two
h
id
d
en
lay
er
s
an
d
a
b
atch
s
ize
o
f
8
is
th
e
b
est
m
o
d
el
f
o
r
th
is
class
.
Fo
r
th
e
Sy
n
o
n
y
m
class
in
T
ab
le
3
,
th
e
r
esu
lts
ac
h
iev
e
th
e
p
er
f
o
r
m
an
ce
u
p
to
9
2
%,
m
o
r
eo
v
e
r
,
in
th
e
s
ec
o
n
d
(
o
n
e
h
id
d
en
la
y
er
)
a
n
d
elev
e
n
th
m
o
d
el
(
th
r
ee
h
id
d
e
n
lay
er
s
)
,
th
e
s
en
s
itiv
ity
g
ets
th
e
p
er
f
ec
t
p
e
r
f
o
r
m
an
ce
,
1
0
0
%.
Dif
f
er
en
t
f
r
o
m
th
r
ee
class
es,
th
e
last
clas
s
o
f
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
,
t
h
e
r
esu
lts
ar
e
n
o
t
g
o
o
d
en
o
u
g
h
,
wi
th
th
e
p
er
ce
n
ta
g
e
u
p
6
8
%.
Fro
m
T
ab
le
4
,
th
e
p
er
f
o
r
m
an
ce
in
s
en
s
itiv
ity
,
p
r
ec
is
io
n
,
an
d
F1
-
sc
o
r
e
ar
e
n
o
t
s
atis
f
y
in
g
.
Ad
d
in
g
m
o
r
e
h
id
d
en
lay
er
s
n
o
t
g
iv
e
s
ig
n
if
ican
t
r
esu
lts
,
b
ec
a
u
s
e
it
d
o
es
n
o
t
r
ea
lly
a
f
f
ec
t
th
e
D
NNs
p
er
f
o
r
m
a
n
ce
.
T
h
e
lim
it
ed
p
ar
a
m
eter
s
th
at
f
in
e
-
tu
n
e
d
(
la
y
er
s
an
d
b
atch
s
ize)
ar
e
n
o
t
alwa
y
s
g
iv
en
a
p
r
o
m
is
in
g
r
esu
lt.
T
h
e
lear
n
in
g
r
a
t
e
an
d
o
p
tim
izer
in
ea
ch
m
o
d
el
ar
e
h
ig
h
ly
t
o
co
n
s
id
er
.
T
h
e
p
r
o
p
o
s
ed
DNNs
ar
ch
i
tectu
r
e
in
ea
ch
class
ca
n
n
o
t
b
e
eq
u
ated
d
u
e
to
th
e
p
r
o
p
o
r
tio
n
o
f
th
e
n
u
m
b
er
o
f
d
ata
is
d
if
f
er
en
t.
T
ab
le
2
.
DNNs
pe
r
f
o
r
m
an
ce
f
o
r
h
o
m
o
n
y
m
class
if
icatio
n
(
class
1)
M
o
d
e
l
P
e
r
f
o
r
ma
n
c
e
E
v
a
l
u
a
t
i
o
n
(
%)
N
u
mb
e
r
o
f
d
a
t
a
A
c
c
u
r
a
c
y
S
e
n
s
i
t
i
v
i
t
y
S
p
e
c
i
f
i
c
i
t
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
1
97
92
98
93
92
1
4
7
2
97
93
98
92
92
1
5
4
3
97
96
98
89
92
1
5
3
4
97
97
97
86
91
1
5
3
5
98
98
99
94
96
1
5
9
6
98
92
99
95
94
1
6
4
7
98
95
98
93
94
1
5
1
8
98
97
98
94
95
1
5
7
9
97
92
98
89
90
1
5
8
10
97
94
98
90
92
1
6
5
11
97
93
97
87
90
1
5
8
12
97
97
98
89
93
1
5
9
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
4
,
Au
g
u
s
t 2
0
2
1
:
1
2
0
8
-
1
2
1
7
1214
T
ab
le
3
.
DNNs
p
er
f
o
r
m
an
ce
f
o
r
s
y
n
o
n
y
m
class
if
icatio
n
(
cla
s
s
2)
M
o
d
e
l
P
e
r
f
o
r
ma
n
c
e
E
v
a
l
u
a
t
i
o
n
(
%)
N
u
mb
e
r
o
f
d
a
t
a
A
c
c
u
r
a
c
y
S
e
n
s
i
t
i
v
i
t
y
S
p
e
c
i
f
i
c
i
t
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
1
98
99
98
95
97
2
1
2
2
98
1
0
0
98
92
96
2
1
8
3
99
99
99
97
98
2
3
2
4
98
99
98
94
96
2
1
8
5
98
99
98
93
96
2
1
5
6
99
99
99
96
97
2
1
0
7
98
98
98
96
97
2
1
8
8
99
99
99
96
97
2
0
4
9
98
99
98
93
96
2
1
9
10
98
97
98
93
95
1
9
6
11
98
1
0
0
98
94
97
2
0
9
12
98
99
98
94
97
2
2
7
T
ab
le
4
.
DNNs
p
er
f
o
r
m
an
ce
f
o
r
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
class
if
i
ca
tio
n
(
class
3)
M
o
d
e
l
P
e
r
f
o
r
ma
n
c
e
E
v
a
l
u
a
t
i
o
n
(
%)
N
u
mb
e
r
o
f
d
a
t
a
A
c
c
u
r
a
c
y
S
e
n
s
i
t
i
v
i
t
y
S
p
e
c
i
f
i
c
i
t
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
1
98
80
99
70
75
2
1
2
2
98
88
99
75
81
2
1
8
3
98
75
99
72
73
2
3
2
4
98
68
99
76
72
2
1
8
5
99
81
99
90
85
2
1
5
6
98
90
98
71
80
2
1
0
7
98
83
98
72
77
2
1
8
8
99
89
99
89
89
2
0
4
9
98
80
99
75
77
2
1
9
10
98
86
99
76
81
1
9
6
11
98
70
99
84
77
2
0
9
12
98
77
99
91
83
2
2
7
I
n
th
is
s
tu
d
y
,
th
e
tr
ain
in
g
d
ata
u
s
e
8
0
%
o
f
th
e
d
ata,
an
d
th
e
r
em
ain
in
g
d
ata
p
e
r
ce
n
tag
e
is
u
s
ed
f
o
r
th
e
test
in
g
p
r
o
ce
s
s
.
T
h
e
twelv
e
m
o
d
els
th
at
im
p
lem
en
ted
in
f
o
u
r
class
o
f
AND
ca
teg
o
r
y
u
s
e
th
e
s
am
e
h
y
p
er
p
ar
am
eter
s
with
t
h
e
lear
n
in
g
r
ate
o
f
0
.
0
0
0
1
.
Fu
r
th
e
r
,
t
h
e
ac
tiv
atio
n
f
u
n
ctio
n
in
th
e
h
id
d
en
lay
e
r
is
R
eL
U
an
d
So
f
tm
a
x
i
n
th
e
o
u
tp
u
t
la
y
e
r
.
T
h
e
n
u
m
b
er
o
f
n
o
d
es
is
1
0
0
in
th
e
in
p
u
t
lay
er
.
T
h
e
to
tal
e
p
o
ch
s
f
o
r
th
e
twelv
e
m
o
d
els
ar
e
5
0
.
T
h
e
co
m
p
a
r
is
o
n
o
f
DNNs
ar
ch
itectu
r
e
in
f
o
u
r
class
es
ca
n
b
e
p
r
esen
te
d
in
T
ab
le
5
.
Fr
o
m
T
ab
le
5
,
th
e
ei
g
h
th
m
o
d
el
s
h
o
wed
ac
cu
r
ac
y
,
s
en
s
itiv
ity
,
s
p
ec
if
icity
,
p
r
ec
is
io
n
,
an
d
F1
-
Sco
r
e,
wh
ich
is
9
8
.
8
5
%,
9
5
.
9
5
%,
9
9
.
2
6
%,
9
4
.
8
0
%,
an
d
9
5
.
3
6
%,
r
esp
ec
tiv
ely
.
Ov
er
all
,
th
e
eig
h
th
m
o
d
el
s
h
o
ws th
e
b
est r
esu
lts
in
o
th
er
m
o
d
els.
T
h
e
p
r
o
p
o
s
ed
DNNs a
r
ch
itec
tu
r
e
is
two
h
id
d
e
n
lay
e
r
s
with
a
b
atch
s
ize
o
f
6
4
.
T
ab
le
5
.
Av
e
r
ag
e
p
e
r
f
o
r
m
an
ce
f
o
r
4
class
es
M
o
d
e
l
P
e
r
f
o
r
ma
n
c
e
E
v
a
l
u
a
t
i
o
n
(
%)
A
c
c
u
r
a
c
y
S
e
n
s
i
t
i
v
i
t
y
S
p
e
c
i
f
i
c
i
t
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
1
9
8
.
3
6
9
2
.
7
4
9
8
.
8
6
8
9
.
5
9
9
0
.
9
2
2
9
8
.
0
1
9
4
.
4
9
9
8
.
6
8
8
9
.
8
0
9
1
.
9
6
3
9
8
.
4
1
9
1
.
9
8
9
9
.
0
1
8
9
.
7
6
9
0
.
8
2
4
9
7
.
8
1
9
0
.
1
7
9
8
.
6
5
8
9
.
1
1
8
9
.
4
7
5
9
8
.
7
1
9
4
.
1
1
9
9
.
1
5
9
4
.
6
5
9
4
.
3
0
6
9
8
.
3
6
9
4
.
9
4
9
8
.
8
6
9
0
.
5
8
9
2
.
4
6
7
9
8
.
4
6
9
3
.
9
0
9
9
.
0
1
9
0
.
3
6
9
2
.
0
1
8
9
8
.
8
5
9
5
.
9
5
9
9
.
2
6
9
4
.
8
0
9
5
.
3
6
9
9
7
.
8
1
9
1
.
9
7
9
8
.
6
1
8
9
.
3
9
9
0
.
6
2
10
9
7
.
9
6
9
3
.
6
7
9
8
.
5
9
9
0
.
0
7
9
1
.
7
7
11
9
7
.
9
6
9
0
.
3
0
9
8
.
5
4
9
1
.
5
5
9
0
.
7
3
12
9
8
.
3
1
9
2
.
7
4
9
8
.
9
2
9
3
.
7
1
9
3
.
0
3
T
o
v
e
r
if
y
th
e
p
r
o
p
o
s
ed
DNN
s
m
o
d
el,
th
e
ac
c
u
r
ac
y
an
d
th
e
lo
s
s
cu
r
v
e
ar
e
s
h
o
wn
in
Fi
g
u
r
es
5
(
a)
an
d
(
b
)
,
r
esp
ec
tiv
ely
.
I
n
Fig
u
r
e
5
(
a
)
,
th
e
ac
cu
r
ac
y
o
f
th
e
tr
ai
n
in
g
an
d
test
in
g
s
et
s
tar
ts
f
r
o
m
ar
o
u
n
d
6
0
%
in
th
e
f
ir
s
t
ep
o
ch
.
L
ater
,
th
e
ac
cu
r
ac
y
is
in
cr
ea
s
ed
with
m
o
r
e
ep
o
ch
s
.
T
h
en
,
in
th
e
ab
o
v
e
1
0
ep
o
c
h
s
,
th
e
ac
cu
r
ac
y
g
ets
th
e
p
er
f
ec
t
s
co
r
e
in
th
e
tr
ai
n
in
g
s
et.
T
h
e
litt
le
g
a
p
b
etwe
en
ac
cu
r
ac
y
i
n
tr
ain
i
n
g
an
d
t
esti
n
g
s
et
lea
d
th
e
o
v
er
f
itti
n
g
.
Fig
u
r
e
5
(
b
)
s
h
o
ws
d
ec
r
ea
s
in
g
er
r
o
r
alo
n
g
with
th
e
in
cr
ea
s
in
g
e
p
o
ch
s
i
n
th
e
tr
ain
in
g
an
d
test
in
g
s
et.
B
o
th
cu
r
v
es p
r
o
d
u
ce
g
o
o
d
s
h
a
p
es in
th
e
p
r
o
ce
s
s
in
g
with
DNN.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
N
eu
r
a
l n
etw
o
r
k
tech
n
iq
u
e
w
ith
d
ee
p
s
tr
u
ctu
r
e
fo
r
imp
r
o
vin
g
a
u
th
o
r
…
(
F
ir
d
a
u
s
)
1215
(
a)
(
b
)
Fig
u
r
e
5
.
Acc
u
r
ac
y
a
n
d
lo
s
s
cu
r
v
e
o
f
p
r
o
p
o
s
ed
d
n
n
s
ar
c
h
itectu
r
e
:
(
a)
ac
cu
r
ac
y
c
u
r
v
e
,
(
b
)
l
o
s
s
cu
r
v
e
I
n
th
is
s
tu
d
y
,
th
e
b
en
ch
m
a
r
k
i
n
g
s
tu
d
y
o
f
DNNs
with
o
th
er
ML
m
eth
o
d
s
is
p
r
o
p
o
s
ed
.
T
h
e
o
th
er
ML
class
if
ier
s
s
u
ch
as
SVM
(
m
atr
i
x
k
er
n
el
p
o
ly
n
o
m
ial,
R
F,
an
d
lin
ea
r
)
,
d
ec
is
io
n
tr
ee
(
C
AR
T
an
d
C
4
.
5
alg
o
r
ith
m
)
,
an
d
co
n
v
e
n
tio
n
al
ar
tific
ial
n
eu
r
al
n
etwo
r
k
s
(
ANNs)
is
co
m
p
a
r
ed
to
th
e
p
r
o
p
o
s
ed
DNNs
ar
c
h
itectu
r
e
with
th
r
ee
h
id
d
en
la
y
er
s
.
T
h
e
r
esu
lt o
f
cl
ass
if
ier
s
ca
n
b
e
s
ee
n
in
T
ab
le
6
,
in
b
o
th
tr
ain
in
g
an
d
test
in
g
s
et.
T
ab
les
6
an
d
7
s
h
o
w
th
e
ac
c
u
r
ac
y
o
f
th
e
n
eu
r
al
n
etw
o
r
k
s
,
wh
eth
er
in
DNNs
o
r
ANNs,
o
u
tp
er
f
o
r
m
SVM
an
d
d
ec
is
io
n
tr
ee
in
all
f
i
v
e
m
etr
ics.
I
t
s
h
o
ws
th
e
n
e
u
r
al
n
etwo
r
k
s
till
b
e
th
e
p
r
o
m
is
in
g
class
if
ier
f
o
r
AND
ca
s
e,
d
u
e
to
n
eu
r
al
n
etwo
r
k
m
u
ch
m
o
r
e
ted
io
u
s
th
a
n
u
s
in
g
a
n
o
f
f
-
th
e
-
s
h
elf
class
if
ier
lik
e
S
VM
.
T
ab
le
7
s
h
o
ws
th
e
s
am
e
ac
cu
r
ac
y
o
f
DNNs
an
d
ANNs
is
9
8
.
8
5
%
i
n
th
e
test
in
g
s
et.
Ho
wev
er
,
f
o
r
th
e
g
r
ea
ter
an
d
a
la
r
g
e
a
m
o
u
n
t
o
f
d
ata
in
th
e
d
ig
ital
lib
r
a
r
y
,
m
ak
in
g
DNNs
m
o
r
e
h
av
e
th
e
p
r
o
s
p
ec
t
o
f
b
ein
g
t
h
e
s
u
g
g
ested
c
lass
if
ier
th
an
ANNs
in
th
e
f
u
tu
r
e.
DNNs
ar
e
m
o
r
e
p
er
f
o
r
m
in
g
with
th
e
co
m
b
in
atio
n
o
f
au
to
m
ated
f
ea
t
u
r
e
r
ep
r
ese
n
tatio
n
in
B
ig
Data
.
T
ab
le
6
.
Acc
u
r
ac
y
b
e
n
ch
m
ar
k
in
g
with
o
th
er
ML
m
eth
o
ds
M
e
t
h
o
d
s
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Te
st
i
n
g
A
c
c
u
r
a
c
y
D
N
N
s (3
l
a
y
e
r
s)
9
9
.
7
3
9
8
.
8
5
S
V
M
w
i
t
h
ma
t
r
i
x
k
e
r
n
e
l
p
o
l
y
n
o
mi
a
l
9
9
.
8
5
9
8
.
5
6
S
V
M
w
i
t
h
ma
t
r
i
x
K
e
r
n
e
l
R
B
F
9
9
.
8
5
9
8
.
6
1
S
V
M
w
i
t
h
M
a
t
r
i
x
K
e
r
n
e
l
Li
n
e
a
r
9
9
.
8
2
9
1
.
8
8
D
e
c
i
s
i
o
n
t
r
e
e
w
i
t
h
C
A
R
T
a
l
g
o
r
i
t
h
m
9
9
.
8
0
9
0
.
2
9
D
e
c
i
s
i
o
n
t
r
e
e
w
i
t
h
C
4
.
5
a
l
g
o
r
i
t
h
m
9
9
.
8
0
8
8
.
9
0
A
N
N
s
9
9
.
6
8
9
8
.
8
5
T
ab
le
7
.
Per
f
o
r
m
an
ce
b
en
c
h
m
ar
k
in
g
with
o
th
er
ML
me
th
o
d
s
M
e
t
h
o
d
s
P
e
r
f
o
r
ma
n
c
e
E
v
a
l
u
a
t
i
o
n
(
%)
A
c
c
u
r
a
c
y
S
e
n
s
i
t
i
v
i
t
y
S
p
e
c
i
f
i
c
i
t
y
P
r
e
c
i
s
i
o
n
F1
-
S
c
o
r
e
DNNs
9
8
.
8
5
9
5
.
9
5
9
9
.
2
6
9
4
.
8
0
9
5
.
3
6
S
V
M
9
8
.
6
1
9
5
.
0
8
9
9
.
1
8
8
9
.
4
4
9
1
.
8
0
D
e
c
i
s
i
o
n
t
r
e
e
9
0
.
2
9
7
4
.
3
4
9
2
.
1
1
6
7
.
9
9
7
0
.
6
4
ANNs
9
8
.
8
0
9
5
.
6
0
9
9
.
2
6
9
4
.
6
1
9
5
.
0
6
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
1
6
9
3
-
6
9
3
0
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
,
Vo
l.
19
,
No
.
4
,
Au
g
u
s
t 2
0
2
1
:
1
2
0
8
-
1
2
1
7
1216
T
h
e
ca
s
e
o
f
AND
is
n
o
t
a
n
e
w
s
tu
d
y
in
b
ib
lio
g
r
ap
h
ic
d
ata.
T
h
is
p
ap
er
also
co
m
p
ar
es
th
e
p
r
o
p
o
s
ed
DNNs
w
ith
s
o
m
e
p
r
ev
io
u
s
lit
er
atu
r
e.
T
ab
le
8
s
h
o
ws
th
e
b
en
ch
m
ar
k
i
n
g
s
tu
d
y
with
co
n
v
e
n
tio
n
al
an
d
n
eu
r
al
n
etwo
r
k
m
eth
o
d
s
.
T
r
an
et
al
.
[
1
9
]
p
r
o
p
o
s
ed
DNNs
f
o
r
au
th
o
r
d
is
am
b
ig
u
atio
n
ca
s
e.
T
h
e
DNNs
m
eth
o
d
ac
h
iev
es
9
9
.
3
1
%
in
ter
m
s
o
f
ac
cu
r
ac
y
.
DNNs
ar
e
s
ig
n
if
ican
tly
o
u
tp
er
f
o
r
m
ed
o
th
er
m
et
h
o
d
s
th
at
u
s
e
a
p
r
ed
ef
in
e
d
f
ea
tu
r
e
s
et.
Ho
wev
er
,
h
o
m
o
n
y
m
s
an
d
s
y
n
o
n
y
m
s
ar
e
ess
en
tial
p
r
o
b
lem
s
in
AND.
Sti
ll,
C
h
in
et
a
l
.
[
2
4
]
im
p
lem
en
t
Mic
r
o
s
o
f
t
Aca
d
em
ic
Sear
ch
(
MA
S)
f
o
r
s
tr
in
g
c
o
m
p
ar
is
o
n
i
n
th
e
au
t
h
o
r
’
s
n
am
e.
T
h
e
m
et
h
o
d
is
m
er
g
in
g
t
h
e
r
esu
lts
o
f
two
p
r
ed
ictio
n
s
f
u
r
t
h
er
b
o
o
s
t th
e
p
er
f
o
r
m
an
ce
.
T
h
is
ap
p
r
o
ac
h
m
et
h
o
d
ac
h
iev
es
F1
-
s
co
r
e
0
.
9
9
2
0
2
o
n
th
e
p
r
iv
ate
lead
er
b
o
ar
d
,
wh
ile
0
.
9
9
1
9
5
o
n
th
e
p
u
b
lic
lead
e
r
b
o
ar
d
.
T
h
en
,
W
an
g
et
a
l
.
[
2
5
]
u
s
e
Pu
b
licatio
n
,
W
eb
Pag
e
&
New
s
Sto
r
ie’
s
d
ataset
with
a
p
air
wis
e
f
ac
to
r
g
r
ap
h
(
PF
G)
an
d
i
n
ter
ac
tio
n
o
f
u
s
er
to
en
h
an
ce
th
e
p
er
f
o
r
m
an
ce
.
T
h
e
F1
-
s
co
r
e
ac
h
iev
es
0
.
8
1
5
in
C
AL
O
d
ataset
ca
s
e.
T
h
e
last
,
with
o
u
r
p
r
o
p
o
s
ed
DNNs
m
eth
o
d
with
th
r
ee
h
id
d
e
n
la
y
er
s
,
DNN
s
ca
n
s
h
o
w
9
8
.
8
5
%
o
f
ac
cu
r
ac
y
in
ca
s
e
o
f
non
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
,
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
,
s
y
n
o
n
y
m
,
a
n
d
h
o
m
o
n
y
m
a
u
th
o
r
s
.
T
ab
le
8
.
Per
f
o
r
m
an
ce
b
en
c
h
m
ar
k
in
g
with
p
r
ev
io
u
s
m
eth
o
d
s
A
u
t
h
o
r
s
M
e
t
h
o
d
D
a
t
a
s
e
t
C
a
se
A
c
c
u
r
a
c
y
F1
-
S
c
o
r
e
Tr
a
n
e
t
a
l
.
[
1
9
]
DNNs
V
i
e
t
n
a
m
e
se
a
u
t
h
o
r
-
n
a
me
A
u
t
h
o
r
a
m
b
i
g
u
a
t
i
o
n
s
9
9
.
3
1
%
-
C
h
i
n
e
t
a
l
.
[
2
4
]
M
e
r
g
i
n
g
r
e
s
u
l
t
s
o
f
t
w
o
p
r
e
d
i
c
t
i
o
n
s
f
u
r
t
h
e
r
b
o
o
st
t
h
e
p
e
r
f
o
r
m
a
n
c
e
M
i
c
r
o
s
o
f
t
A
c
a
d
e
mi
c
S
e
a
r
c
h
(
M
A
S
)
S
t
r
i
n
g
C
o
mp
a
r
i
s
o
n
-
0
.
9
9
W
a
n
g
e
t
a
l
.
[
2
5
]
A
p
a
i
r
w
i
se
f
a
c
t
o
r
g
r
a
p
h
(
P
F
G
)
mo
d
e
l
P
u
b
l
i
c
a
t
i
o
n
,
W
e
b
P
a
g
e
&
N
e
w
s St
o
r
i
e
s
C
A
LO
D
a
t
a
se
t
-
0
.
8
2
P
r
o
p
o
se
d
DNNs
D
B
LP
N
o
n
-
H
o
mo
n
y
m
-
S
y
n
o
n
y
m,
H
o
m
o
n
y
m
-
S
y
n
o
n
y
m
,
S
y
n
o
n
y
m
,
a
n
d
H
o
mo
n
y
m
A
u
t
h
o
r
s
9
8
.
8
5
%
0
.
9
5
4.
CO
NCLU
SI
O
N
T
h
e
m
ain
ch
allen
g
e
to
AND
i
s
th
at
th
e
d
ata
is
b
ec
o
m
in
g
m
o
r
e
an
d
m
o
r
e
co
m
p
le
x
a
n
d
d
y
n
am
ic.
I
t
r
eq
u
ir
es
th
e
AND
alg
o
r
ith
m
t
o
b
e
ex
te
n
s
ib
le
an
d
f
lex
ib
le
f
o
r
d
if
f
e
r
en
t
s
ce
n
ar
io
s
.
T
h
is
p
a
p
er
p
r
o
p
o
s
es
o
n
e
o
f
ML
alg
o
r
ith
m
s
,
DNNs,
f
o
r
m
u
lticlas
s
cla
s
s
if
icat
io
n
in
non
-
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
,
h
o
m
o
n
y
m
-
s
y
n
o
n
y
m
,
s
y
n
o
n
y
m
,
an
d
h
o
m
o
n
y
m
au
th
o
r
s
task
.
Fo
r
th
e
class
if
icatio
n
task
,
a
k
n
o
wled
g
e
b
ase
is
r
eq
u
ir
ed
.
T
h
e
DB
L
P
co
m
p
u
ter
s
cien
ce
b
ib
lio
g
r
ap
h
y
d
ataset
with
f
iv
e
attr
ib
u
tes
h
av
e
im
p
lem
en
ted
f
o
r
th
e
AND
p
r
o
b
lem
s
.
E
x
p
er
im
en
tal
r
esu
lts
s
h
o
w
th
at
o
n
all
th
e
twelv
e
d
if
f
er
en
t
m
o
d
els,
o
u
r
p
r
o
p
o
s
ed
m
eth
o
d
o
u
tp
er
f
o
r
m
s
th
e
co
m
m
o
n
ex
is
tin
g
ML
m
et
h
o
d
s
,
s
u
ch
as
SVM
a
n
d
d
ec
is
io
n
tr
ee
.
T
h
e
DNNs
p
e
r
f
o
r
m
an
ce
with
th
r
ee
h
i
d
d
en
lay
er
s
as
th
e
b
est
m
o
d
el,
ac
h
iev
e
ac
cu
r
ac
y
,
s
en
s
itiv
ity
,
s
p
ec
if
icity
,
p
r
ec
is
io
n
,
an
d
F1
-
s
co
r
e
a
r
e
9
8
.
8
5
%,
9
5
.
9
5
%,
9
9
.
2
6
%,
9
4
.
8
0
%,
an
d
9
5
.
3
6
%,
r
esp
ec
tiv
ely
.
I
n
th
e
f
u
tu
r
e
,
DNNs
ar
e
m
o
r
e
p
er
f
o
r
m
in
g
with
th
e
au
to
m
ated
f
ea
tu
r
e
r
ep
r
esen
tatio
n
in
AN
D
p
r
o
ce
s
s
in
g
.
RE
F
E
R
E
NC
E
S
[1
]
K.
Ca
n
e
se
a
n
d
S
.
Weis,
“
P
u
b
M
e
d
:
th
e
b
ib
l
io
g
ra
p
h
ic
d
a
tab
a
se
,
”
Na
t
io
n
a
l
L
ib
ra
ry
o
f
M
e
d
icin
e
,
2
0
1
3
.
[2
]
J.
Kim
a
n
d
J.
Kim
,
“
Th
e
imp
a
c
t
o
f
imb
a
lan
c
e
d
train
in
g
d
a
ta
o
n
m
a
c
h
in
e
lea
rn
in
g
fo
r
a
u
t
h
o
r
n
a
m
e
d
i
sa
m
b
ig
u
a
ti
o
n
,
”
S
c
ien
to
me
trics
,
v
o
l.
1
1
7
,
n
o
.
1
,
p
p
.
5
1
1
-
5
2
6
,
A
u
g
.
2
0
1
8
,
d
o
i
:
1
0
.
1
0
0
7
/s1
1
1
9
2
-
0
1
8
-
2
8
6
5
-
9.
[3
]
C.
A.
D
.
An
g
e
lo
,
C
.
G
iu
ffri
d
a
,
G
.
Ab
ra
m
o
,
a
n
d
C.
A.
D.
An
g
e
lo
,
“
A
h
e
u
risti
c
a
p
p
r
o
a
c
h
to
a
u
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
t
io
n
in
b
ib
li
o
m
e
tri
c
s
d
a
tab
a
se
s
fo
r
larg
e
-
sc
a
le
re
se
a
rc
h
a
ss
e
s
sm
e
n
ts,”
J
.
Am.
S
o
c
.
In
f.
S
c
i.
T
e
c
h
n
o
l.
,
v
o
l.
6
2
,
n
o
.
2
,
p
p
.
2
5
7
-
2
6
9
,
2
0
1
1
,
d
o
i:
1
0
.
1
0
0
2
/a
si
.
2
1
4
6
0
.
[4
]
K.
-
H.
Ya
n
g
,
H.
-
T
.
P
e
n
g
,
J.
-
Y.
Jia
n
g
,
H.
-
M
.
Lee
,
a
n
d
J.
-
M
.
Ho
,
“
Au
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
ti
o
n
f
o
r
c
it
a
ti
o
n
s
u
si
n
g
to
p
ic
a
n
d
we
b
c
o
rre
latio
n
,
”
Res
.
Ad
v
.
T
e
c
h
n
o
l.
Di
g
it
.
L
i
b
r.
,
p
p
.
1
8
5
–
1
9
6
,
S
e
p
.
2
0
0
8
,
d
o
i:
1
0
.
1
0
0
7
/9
7
8
-
3
-
5
4
0
-
8
7
5
9
9
-
4
_
1
9
.
[5
]
A.
A.
F
e
rre
ira,
A.
Ve
lo
so
,
M
.
A.
G
o
n
ç
a
lv
e
s,
a
n
d
A.
H.
F
.
Lae
n
d
e
r,
“
Eff
e
c
ti
v
e
se
lf
-
train
i
n
g
a
u
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
t
io
n
i
n
sc
h
o
larly
d
i
g
it
a
l
li
b
ra
ries
,
”
Pr
o
c
e
e
d
in
g
s
o
f
th
e
1
0
th
a
n
n
u
a
l
jo
in
t
c
o
n
fer
e
n
c
e
o
n
Dig
it
a
l
li
b
ra
rie
s
,
2
0
1
0
,
p
p
.
3
9
-
48
,
d
o
i:
1
0
.
1
1
4
5
/
1
8
1
6
1
2
3
.
1
8
1
6
1
3
0
.
[6
]
V.
I.
T.
a
n
d
N.
R
.
S
m
a
lh
e
ise
r,
“
NIH
P
u
b
li
c
Ac
c
e
ss
,
”
US
N
a
ti
o
n
a
l
L
ib
ra
ry
o
f
M
e
d
ici
n
e
,
v
o
l
.
3
,
n
o
.
3
,
p
p
.
1
–
3
0
,
2
0
1
0
.
[7
]
M
.
R.
Ac
k
e
rm
a
n
n
a
n
d
F
.
Re
it
z
,
“
Ho
m
o
n
y
m
d
e
tec
ti
o
n
in
c
u
ra
ted
b
ib
li
o
g
ra
p
h
ies
:
lea
rn
in
g
fr
o
m
d
b
l
p
’s
e
x
p
e
rien
c
e
,
”
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
T
h
e
o
ry
a
n
d
Pra
c
ti
c
e
o
f
Dig
it
a
l
L
ib
r
a
ri
e
s
,
Ju
n
.
2
0
1
8
,
p
p
.
5
9
-
6
5
.
[8
]
J.
Kim
,
“
Ev
a
lu
a
ti
n
g
a
u
t
h
o
r
n
a
m
e
d
isa
m
b
ig
u
a
ti
o
n
fo
r
d
ig
it
a
l
li
b
ra
r
ies
:
a
c
a
se
o
f
DBLP
,
”
S
c
ien
to
me
trics
,
v
o
l.
1
1
6
,
n
o
.
3
,
p
p
.
1
8
6
7
-
1
8
8
6
,
2
0
1
8
.
[9
]
F
.
M
o
m
e
n
i
a
n
d
P
.
M
a
y
r,
“
Us
in
g
co
-
a
u
th
o
rsh
i
p
n
e
two
rk
s
fo
r
a
u
t
h
o
r
n
a
m
e
d
isa
m
b
i
g
u
a
ti
o
n
,
”
Pro
c
.
1
6
th
ACM
/IE
EE
-
CS
J
t.
C
o
n
f.
Dig
i
t.
L
i
b
r.
-
J
CDL
’
1
6
,
p
p
.
2
6
1
-
2
6
2
,
Ju
n
.
2
0
1
6
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KOM
NI
KA
T
elec
o
m
m
u
n
C
o
m
p
u
t E
l Co
n
tr
o
l
N
eu
r
a
l n
etw
o
r
k
tech
n
iq
u
e
w
ith
d
ee
p
s
tr
u
ctu
r
e
fo
r
imp
r
o
vin
g
a
u
th
o
r
…
(
F
ir
d
a
u
s
)
1217
[1
0
]
F
.
M
o
m
e
n
i
a
n
d
P
.
M
a
y
r,
“
E
v
a
lu
a
ti
n
g
c
o
-
a
u
t
h
o
rsh
ip
n
e
two
rk
s
in
a
u
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
ti
o
n
fo
r
c
o
m
m
o
n
n
a
m
e
s,”
In
ter
n
a
t
io
n
a
l
C
o
n
fer
e
n
c
e
o
n
T
h
e
o
ry
a
n
d
Pra
c
ti
c
e
o
f
Dig
it
a
l
L
ib
r
a
ri
e
s
,
2
0
1
6
,
p
p
.
3
8
6
-
3
9
1
.
[1
1
]
J.
Kim
a
n
d
J.
Die
sn
e
r,
“
Disto
r
ti
v
e
e
ffe
c
ts
o
f
in
it
ial
-
b
a
se
d
n
a
m
e
d
isa
m
b
ig
u
a
t
io
n
o
n
m
e
a
su
re
m
e
n
ts
o
f
larg
e
-
sc
a
le
c
o
a
u
th
o
rsh
i
p
n
e
tw
o
rk
s,”
J
.
Ass
o
c
.
In
f.
S
c
i.
T
e
c
h
n
o
l.
,
v
o
l.
6
7
,
n
o
.
6
,
p
p
.
1
4
4
6
-
1
4
6
1
,
2
0
1
6
,
d
o
i:
1
0
.
1
0
0
2
/as
i.
2
3
4
8
9
.
[1
2
]
S
.
He
e
ffe
r,
B.
T
h
ij
s,
a
n
d
W.
G
län
z
e
l,
“
Are
re
g
istere
d
a
u
th
o
r
s
m
o
re
p
r
o
d
u
c
ti
v
e
?
,
”
Pro
c
.
IS
S
I
2
0
1
3
,
v
o
l
.
2
,
p
p
.
1
8
6
4
-
1
8
6
7
,
Ja
n
u
a
ry
2
0
1
3
.
[1
3
]
D.
E.
S
m
it
h
,
J.
F
ra
n
k
,
a
n
d
W.
Cu
s
h
in
g
,
“
Th
e
AN
M
L
lan
g
u
a
g
e
,
”
T
h
e
ICAP
S
-
0
8
W
o
rk
sh
o
p
o
n
Kn
o
wled
g
e
En
g
i
n
e
e
rin
g
fo
r P
l
a
n
n
in
g
a
n
d
S
c
h
e
d
u
li
n
g
(KE
PS
)
,
2
0
0
8
.
[1
4
]
S
.
G
o
ld
m
a
n
a
n
d
Y.
Z
h
o
u
,
“
E
n
h
a
n
c
in
g
su
p
e
rv
ise
d
lea
rn
in
g
wit
h
u
n
l
a
b
e
led
d
a
ta,”
ICM
L
,
2
0
0
0
,
p
p
.
3
2
7
-
3
3
4
.
[1
5
]
Q.
V
Le,
“
Bu
i
ld
i
n
g
h
i
g
h
-
le
v
e
l
fe
a
tu
re
s
u
sin
g
lar
g
e
sc
a
le
u
n
su
p
e
rv
ise
d
lea
rn
i
n
g
,
”
2
0
1
3
IEE
E
in
ter
n
a
ti
o
n
a
l
c
o
n
fer
e
n
c
e
o
n
a
c
o
u
st
ics
,
sp
e
e
c
h
a
n
d
sig
n
a
l
p
ro
c
e
ss
in
g
,
De
c
.
2
0
1
3
,
p
p
.
8
5
9
5
-
8
5
9
8
.
[1
6
]
J.
D.
M
.
Re
n
n
ie
a
n
d
R.
Rif
k
i
n
,
“
Im
p
ro
v
i
n
g
m
u
lt
icla
ss
tex
t
c
las
sifica
ti
o
n
with
th
e
su
p
p
o
rt
v
e
c
to
r
m
a
c
h
in
e
,
”
M
a
ss
a
c
h
u
se
tt
s
In
stit
u
te
o
f
Tec
h
n
o
lo
g
y
a
s
AI
M
e
m
o
2
0
0
1
-
0
2
6
a
n
d
C
BCL
M
e
m
o
2
1
0
,
Oc
t.
2
0
0
1
.
[On
li
n
e
]
.
Av
a
il
a
b
le
:
h
tt
p
:
//
h
d
l.
h
a
n
d
le.n
e
t/
1
7
2
1
.
1
/7
2
4
1
[1
7
]
G
.
M
.
F
o
o
d
y
a
n
d
A.
M
a
th
u
r,
“
A
re
lativ
e
e
v
a
lu
a
ti
o
n
o
f
m
u
l
ti
c
las
s
ima
g
e
c
las
sifica
ti
o
n
b
y
su
p
p
o
rt
v
e
c
to
r
m
a
c
h
in
e
s,”
IEE
E
T
ra
n
s.
Ge
o
sc
i.
Rem
o
te S
e
n
s
.
,
v
o
l.
4
2
,
n
o
.
6
,
p
p
.
1
3
3
5
-
1
3
4
3
,
J
u
n
.
2
0
0
4
,
d
o
i:
1
0
.
1
1
0
9
/T
G
RS
.
2
0
0
4
.
8
2
7
2
5
7
.
[1
8
]
W.
Li
u
,
e
t
a
l.
,
“
A
su
r
v
e
y
o
f
d
e
e
p
n
e
u
ra
l
n
e
two
r
k
a
rc
h
it
e
c
tu
re
s
a
n
d
th
e
ir
a
p
p
li
c
a
ti
o
n
s,”
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l.
2
3
4
,
p
p
.
1
1
-
2
6
,
Ap
r
.
2
0
1
7
,
d
o
i
:
1
0
.
1
0
1
6
/j
.
n
e
u
c
o
m
.
2
0
1
6
.
1
2
.
0
3
8
.
[1
9
]
H.
N.
Tran
,
T.
Hu
y
n
h
,
a
n
d
T.
Do
,
“
Au
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
t
io
n
b
y
u
si
n
g
d
e
e
p
n
e
u
ra
l
n
e
two
rk
,
”
Asia
n
C
o
n
fer
e
n
c
e
o
n
I
n
telli
g
e
n
t
I
n
fo
rm
a
ti
o
n
a
n
d
Da
ta
b
a
se
S
y
ste
ms
,
Ap
r.
2
0
1
4
,
p
p
.
1
2
3
-
1
3
2
.
[2
0
]
H.
Ha
n
,
L.
G
il
e
s,
H.
Zh
a
,
C.
Li
,
a
n
d
K.
Tsi
o
u
tsio
u
li
k
li
s,
“
Two
su
p
e
r
v
ise
d
lea
rn
in
g
a
p
p
ro
a
c
h
e
s
fo
r
n
a
m
e
d
isa
m
b
ig
u
a
t
io
n
in
a
u
t
h
o
r
c
it
a
ti
o
n
s
,
”
Pro
c
e
e
d
in
g
s
o
f
t
h
e
2
0
0
4
J
o
i
n
t
A
C
M
/IE
EE
Co
n
fer
e
n
c
e
o
n
Dig
i
ta
l
L
ib
ra
rie
s
,
2
0
0
4
,
p
p
.
2
9
6
-
3
0
5
.
[2
1
]
C.
S
.
a
n
d
G
.
M
a
rc
h
io
n
i
n
i,
“
Aw
a
re
n
e
ss
in
CIS
,
”
In
t.
Rev
.
Res
.
Op
e
n
Dista
n
c
e
L
e
a
rn
,
v
o
l
.
1
4
,
n
o
.
4
,
p
p
.
9
0
–
1
0
3
,
2
0
1
3
.
[2
2
]
A.
F
.
S
a
n
tan
a
,
M
.
A.
G
o
n
ç
a
lv
e
s,
A.
H.
F
.
Lae
n
d
e
r,
a
n
d
A.
A.
F
e
rre
ira,
“
On
th
e
c
o
m
b
i
n
a
ti
o
n
o
f
d
o
m
a
in
-
sp
e
c
ifi
c
h
e
u
risti
c
s
fo
r
a
u
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
ti
o
n
:
th
e
n
e
a
re
st
c
lu
ste
r
m
e
th
o
d
,
”
I
n
t.
J
.
Di
g
it
.
L
ib
r.
,
v
o
l.
1
6
,
n
o
.
3
,
p
p
.
2
2
9
-
2
4
6
,
2
0
1
5
.
[2
3
]
M.
C.
M
ü
ll
e
r,
F
.
Re
it
z
,
a
n
d
N.
R
o
y
,
“
Da
ta
se
ts
fo
r
a
u
th
o
r
n
a
m
e
d
isa
m
b
ig
u
a
ti
o
n
:
a
n
e
m
p
iri
c
a
l
a
n
a
l
y
sis
a
n
d
a
n
e
w
re
so
u
rc
e
,
”
S
c
ien
to
me
trics
,
v
o
l
.
1
1
1
,
n
o
.
3
,
p
p
.
1
4
6
7
-
1
5
0
0
,
M
a
r.
2
0
1
7
.
[2
4
]
W.
-
S
.
C
h
i
n
,
Y.
Ch
i
n
J
u
a
n
,
Y.
Z
h
u
a
n
g
,
F
.
Wu
,
“
Eff
e
c
ti
v
e
stri
n
g
p
r
o
c
e
ss
in
g
a
n
d
m
a
tch
in
g
f
o
r
a
u
th
o
r
d
i
sa
m
b
ig
u
a
ti
o
n
,
”
J
.
M
a
c
h
.
L
e
a
r
n
.
Res
.
,
v
o
l.
1
5
,
n
o
.
1
,
p
p
.
3
0
3
7
-
3
0
6
4
,
Au
g
.
2
0
1
4
,
d
o
i:
1
0
.
1
1
4
5
/2
5
1
7
2
8
8
.
2
5
1
7
2
9
5
.
[2
5
]
X.
Wan
g
,
J.
Tan
g
,
H.
C
h
e
n
g
,
a
n
d
S
.
Y.
P
h
il
i
p
,
“
A
d
a
n
a
:
Ac
ti
v
e
n
a
m
e
d
isa
m
b
ig
u
a
ti
o
n
,
”
D
a
ta
M
in
i
n
g
(ICD
M
),
2
0
1
1
IEE
E
1
1
th
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
,
De
c
.
2
0
1
1
,
p
p
.
7
9
4
-
8
0
3
,
d
o
i:
1
0
.
1
1
0
9
/ICD
M
.
2
0
1
1
.
1
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.