I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
A
I
)
V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024, pp.
4323
~
4333
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v13.i4.pp
4323
-
4333
4323
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
Im
p
r
ovi
n
g t
h
e
r
i
sk
p
r
of
i
l
e
of
I
n
d
on
e
s
i
an
e
n
t
e
r
p
r
i
se
t
axp
aye
r
s
u
si
n
g
m
u
l
t
i
l
ab
e
l
c
l
ass
i
f
i
c
at
i
on
T
e
gu
h
P
r
as
e
t
yo, B
u
d
i
S
u
s
e
t
yo, An
an
g K
u
r
n
ia
D
e
pa
r
t
m
e
nt
of
S
t
a
t
i
s
t
i
c
s
, F
a
c
ul
t
y of
M
a
t
he
m
a
t
i
c
s
a
nd
N
a
t
ur
a
l
S
c
i
e
nc
e
s
, I
P
B
U
n
i
ve
r
s
i
t
y, B
ogor
,
I
ndone
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
N
ov 28, 2023
R
e
vi
s
e
d
F
e
b 29, 2024
A
c
c
e
pt
e
d
M
a
r
21, 2024
Optimizing
tax
revenues
is
difficult
in
Indonesia
due
to
obstacles
such
as
tax
evasion
and
tax
avoidance.
It
is
closely
related
to
an
organizati
on'
s
complian
ce
with
tax
regulations,
known
as
the
taxpayer
s
risk
profile
.
However,
this
mechanism
does
not
accura
tely
detect
tax
avoidance
and
tax
evasion
ris
ks.
To
overcome
this
limitation,
we
use
a
multilabel
classification
machine
le
arning
method
in
this
study,
which
classifies
a
single
observation
into
one
o
r
more
labels
a
t
once.
The
approach
involves
problem
transformation
(
b
inary
relevance
and
label
powerset
),
algorit
hm
adaptatio
n
(
multilabel
k
-
nearest
neighbo
r
(
ML
-
kNN
)
and
multilabel
-
adaptive
resonance
associati
ve
ma
p
(
ML
-
ARAM
)
)
,
and
e
nsemble
(
l
abel
space
partitioning
and
r
andom
k
-
label
se
ts
with
disjoint
(RAkELd)
).
Based
on
the
model
performance
compariso
ns,
we
discovered
that
the
ML
-
ARAM
method
based
on
deep
learning
is
th
e
best,
with
an
average
F1
-
s
core
of
95.5%
and
a
h
amming
l
oss
of
7.4%.
We
also
examine
the
feature
import
ance
of
the
best
model
to
reduce
the
dimensi
ons
of
features
so
that
we
can
identify
the
dominant
factors
that
encourage
a
ta
xpayer
entity
to
engage
in
tax
avoidance
or
tax
evasion.
The
findings
of
this
study
i
mprove
the
accuracy
o
f
tax
avoidance
risk
detection
and
tax
evasion
risk
profiles
using machine learning methods, ensuring maximum tax
reve
nues in
Indonesia
.
K
e
y
w
o
r
ds
:
D
e
e
p l
e
a
r
ni
ng
F
e
a
tu
r
e
i
m
por
ta
nc
e
M
a
c
hi
ne
l
e
a
r
ni
ng
M
ul
ti
la
be
l
c
la
s
s
if
ic
a
ti
on
T
a
x r
e
ve
nue
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
T
e
guh P
r
a
s
e
ty
o
D
e
pa
r
tm
e
nt
of
S
ta
ti
s
ti
c
s
, F
a
c
ul
ty
of
M
a
th
e
m
a
ti
c
s
a
nd
N
a
tu
r
a
l
S
c
ie
nc
e
s
, I
P
B
U
ni
ve
r
s
it
y
B
ogor
, I
ndone
s
ia
E
m
a
il
:
te
guhpr
a
s
e
ty
o@
a
pps
.i
pb.a
c
.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
T
a
xa
ti
on
is
I
ndone
s
ia
’
s
m
os
t
s
ig
ni
f
ic
a
nt
s
ta
te
r
e
ve
nu
e
,
w
hi
c
h
is
c
r
it
ic
a
l
f
or
e
c
onomi
c
de
ve
lo
pm
e
nt
.
I
n
pr
a
c
ti
c
e
, t
a
xpa
y
e
r
s
e
ng
a
ge
i
n
t
a
x
a
voi
da
nc
e
a
nd t
a
x
e
va
s
io
n e
f
f
or
ts
t
o a
voi
d or
e
ve
n
f
a
il
t
o pa
y
t
a
x obli
ga
ti
on
s
,
r
e
s
ul
ti
ng
in
s
ubopti
m
a
l
ta
x
r
e
ve
nue
s
in
I
ndone
s
ia
[
1]
.
A
c
c
or
di
n
g
to
S
in
a
ga
e
t
al
.
[
2]
,
ta
x
a
voi
da
nc
e
oc
c
ur
s
in
I
ndone
s
ia
by
r
e
duc
in
g
pr
of
it
m
a
r
gi
ns
to
a
voi
d
a
ppl
ic
a
bl
e
ta
x
r
a
te
s
or
by
e
xpl
oi
ti
ng
le
ga
l
lo
ophole
s
to
r
e
duc
e
ta
x
li
a
bi
li
ti
e
s
.
M
e
a
nw
hi
le
,
ta
x
e
va
s
io
n,
a
c
c
or
di
ng
to
P
os
te
a
[
3]
,
is
us
e
d
by
de
c
e
pt
io
n
th
a
t
not
le
ga
l
.
T
he
I
ndone
s
ia
n
gove
r
nm
e
nt
is
a
tt
e
m
pt
in
g
to
r
e
duc
e
ta
x
a
voi
da
nc
e
a
n
d
e
va
s
io
n
in
s
e
ve
r
a
l
w
a
ys
,
in
c
lu
di
ng
in
c
r
e
a
s
in
g
ta
x
c
om
pl
ia
nc
e
th
r
ough
th
e
ta
xpa
ye
r
r
is
k
pr
o
f
il
e
.
T
he
m
or
e
c
om
pl
ia
nt
a
n
e
nt
r
e
pr
e
ne
ur
is
,
th
e
gr
e
a
te
r
th
e
oppor
tu
ni
ty
f
o
r
ta
x
r
e
ve
nue
to
be
ge
ne
r
a
te
d.
T
he
obs
ta
c
le
in
th
e
f
ie
ld
is
th
a
t
th
e
c
ur
r
e
nt
r
is
k
pr
o
f
il
e
m
e
c
ha
ni
s
m
f
a
il
s
to
a
c
c
ur
a
te
ly
d
e
te
c
t
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
v
a
s
io
n
r
is
k
s
,
m
a
ki
ng
it
in
e
f
f
e
c
ti
ve
a
nd
in
e
f
f
ic
ie
nt
to
im
pl
e
m
e
nt
.
T
hi
s
l
e
a
d
s
t
o
s
ub
opt
im
a
l
ta
x r
e
ve
nue
i
n I
ndone
s
ia
.
S
e
ve
r
a
l
pr
e
vi
ous
s
tu
di
e
s
us
e
d
m
a
c
hi
ne
le
a
r
ni
ng
m
e
th
od
s
[
4]
–
[
6]
to
id
e
nt
if
y
th
e
r
is
k
of
ta
x a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n. T
hos
e
pr
e
vi
ous
s
tu
di
e
s
e
x
a
m
in
e
th
e
us
e
of
c
la
s
s
if
ic
a
ti
on
m
e
th
ods
in
de
te
c
ti
ng
ta
x
a
voi
d
a
nc
e
a
nd
ta
x
e
va
s
io
n
r
is
ks
,
a
s
w
e
ll
a
s
pr
e
di
c
t
r
e
gi
ona
l
de
ve
lo
pm
e
nt
pr
oj
e
c
t
pol
ic
ie
s
in
s
e
ve
r
a
l
c
ount
r
ie
s
us
in
g
bi
na
r
y
-
c
la
s
s
c
l
a
s
s
if
ic
a
ti
on
(
B
C
C
)
m
e
th
od.
H
ow
e
v
e
r
,
th
e
r
is
k
s
of
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
in
I
ndone
s
ia
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4323
-
4333
4324
a
r
e
di
ve
r
s
e
,
w
it
h
m
or
e
th
a
n
two
la
be
l
c
a
te
gor
ie
s
.
F
ur
th
e
r
m
or
e
, be
c
a
us
e
a
s
in
gl
e
ta
xpa
ye
r
e
nt
it
y c
a
n
e
ng
a
ge
in
one
or
m
or
e
t
ype
s
of
t
a
x a
voi
da
nc
e
or
t
a
x e
va
s
io
n s
im
ul
ta
ne
ous
ly
,
B
C
C
m
ode
li
ng c
a
nnot
be
us
e
d. T
o a
ddr
e
s
s
th
is
li
m
it
a
ti
on,
w
e
pr
e
s
e
nt
a
m
ode
li
ng
s
ol
ut
io
n
in
th
is
s
tu
dy
th
a
t
e
m
pl
oys
th
e
m
ul
ti
la
be
l
c
la
s
s
if
ic
a
ti
on
(
M
L
C
)
m
a
c
hi
ne
le
a
r
ni
ng
m
e
th
od,
w
hi
c
h
c
a
n
m
ode
l
two
or
m
or
e
la
be
l
c
la
s
s
e
s
s
im
ul
ta
ne
ou
s
ly
.
T
hi
s
s
tu
dy
not
onl
y
imp
r
ove
s
th
e
r
is
k
pr
of
il
e
f
or
de
te
c
ti
ng
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
in
I
ndone
s
ia
,
but
it
a
ls
o
in
ve
s
ti
ga
t
e
s
th
e
dom
in
a
nt
f
a
c
to
r
s
t
ha
t
e
nc
our
a
ge
a
t
a
xpa
y
e
r
e
nt
it
y t
o e
nga
ge
i
n t
a
x a
voi
da
nc
e
or
t
a
x e
va
s
io
n.
T
he
c
la
s
s
if
ic
a
ti
on
m
e
th
od
is
us
e
d
to
m
ode
l
r
e
s
pon
s
e
va
r
ia
bl
e
s
a
s
la
be
l
c
la
s
s
c
a
te
gor
ie
s
,
w
hi
c
h
a
r
e
w
id
e
ly
us
e
d
in
va
r
io
us
f
ie
ld
s
,
in
c
lu
di
ng
c
us
to
m
e
r
ta
r
ge
t
m
a
r
ke
ti
ng,
di
s
e
a
s
e
di
a
gnos
i
s
,
doc
um
e
nt
c
la
s
s
if
ic
a
ti
on,
a
nd
s
oc
ia
l
ne
twor
k
a
n
a
ly
s
is
[
7]
.
O
ne
obs
e
r
ve
d
va
lu
e
is
ty
pi
c
a
ll
y
c
la
s
s
if
ie
d
in
to
on
e
c
l
a
s
s
or
la
be
l
on
th
e
r
e
s
pons
e
va
r
ia
bl
e
,
c
a
ll
e
d
B
C
C
or
m
ul
ti
-
c
la
s
s
c
la
s
s
if
ic
a
ti
on
(
M
C
C
)
.
T
he
M
L
C
m
e
th
od
a
s
s
o
c
ia
te
s
a
s
in
gl
e
obs
e
r
va
ti
on
w
it
h
m
ul
ti
pl
e
la
be
ls
[
8]
.
S
e
ve
r
a
l
a
ppl
ic
a
ti
ons
of
th
e
M
L
C
m
e
th
od,
a
m
ong
ot
he
r
s
,
w
e
r
e
c
onduc
te
d
in
[
9]
–
[
12]
,
w
hi
c
h
w
e
r
e
li
m
it
e
d
onl
y
to
m
ode
li
ng
a
nd
m
e
th
od
c
om
pa
r
is
on.
T
hi
s
s
tu
dy
ha
s
ye
t
to
e
xa
m
in
e
da
ta
li
m
it
a
ti
ons
s
uc
h
a
s
c
l
a
s
s
im
ba
la
nc
e
in
th
e
M
L
C
m
od
e
l
a
nd
how
to
a
ddr
e
s
s
it
.
I
n
f
ie
ld
c
ondi
ti
ons
,
d
a
ta
in
di
c
a
ti
ng
c
la
s
s
im
ba
la
nc
e
a
r
e
f
r
e
que
nt
ly
f
ound,
a
s
e
vi
de
nc
e
d
by
th
is
s
tu
dy'
s
e
m
pi
r
ic
a
l
f
in
di
ngs
.
I
n
a
ddi
ti
on
to
c
om
pa
r
in
g
m
e
th
ods
a
nd
id
e
nt
if
yi
ng
r
is
k
f
a
c
to
r
s
f
or
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n,
th
is
s
tu
dy
di
s
c
us
s
e
s
how
to
de
a
l
w
it
h
c
la
s
s
i
m
ba
la
n
c
e
, w
hi
c
h c
a
n he
lp
w
it
h t
he
i
m
pl
e
m
e
nt
a
ti
on of
t
he
M
L
C
m
ode
l
in
r
e
a
l
-
w
or
ld
c
a
s
e
s
.
W
e
us
e
e
m
pi
r
ic
a
l
da
ta
f
r
om
th
e
I
ndone
s
ia
n
M
in
i
s
tr
y
of
F
in
a
nc
e
on
hi
s
to
r
ic
a
l
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
da
ta
f
r
om
2018
to
2022.
I
n
th
is
s
tu
dy
,
w
e
c
om
pa
r
e
s
ix
M
L
C
m
e
th
ods
to
de
te
r
m
in
e
w
hi
c
h
m
ode
l
pe
r
f
or
m
s
th
e
be
s
t
ba
s
e
d
on
F
1
-
s
c
or
e
a
nd
ha
m
m
in
g
lo
s
s
.
U
s
in
g
th
e
be
s
t
m
ode
l,
w
e
id
e
nt
if
y
th
e
r
is
k
e
le
m
e
nt
f
a
c
to
r
s
th
a
t
e
nc
our
a
ge
a
ta
xpa
ye
r
e
nt
it
y
to
e
nga
ge
in
ta
x
a
v
oi
da
nc
e
or
ta
x
e
va
s
io
n
ba
s
e
d
on
th
e
f
e
a
tu
r
e
im
por
ta
nc
e
va
lu
e
.
N
e
xt
,
w
e
us
e
th
e
s
e
dom
in
a
nt
f
a
c
to
r
s
to
r
e
duc
e
r
is
k
e
le
m
e
nt
s
,
m
a
ki
ng
th
e
m
e
a
s
ie
r
to
in
te
r
pr
e
t
a
nd de
c
id
e
. A
c
c
or
di
ng t
o t
hi
s
e
xpl
a
na
ti
on, us
in
g t
he
M
L
C
m
a
c
hi
ne
l
e
a
r
ni
ng me
th
od i
n t
hi
s
s
tu
dy w
il
l
pr
ovi
de
a
s
ol
ut
io
n
to
im
p
r
ove
th
e
a
c
c
ur
a
c
y
of
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
v
a
s
io
n
r
is
k
de
te
c
ti
on,
r
e
s
ul
ti
ng
in
in
c
r
e
a
s
e
d
ta
x
r
e
ve
nue
f
or
I
ndone
s
ia
.
T
he
a
dde
d
va
lu
e
i
s
th
a
t
th
is
r
e
s
e
a
r
c
h
w
il
l
r
e
ve
a
l
th
e
dom
in
a
nt
f
a
c
to
r
s
th
a
t
e
nc
our
a
ge
a
ta
xpa
ye
r
e
nt
it
y
to
e
nga
ge
in
ta
x
a
voi
da
nc
e
or
ta
x
e
va
s
io
n.
I
t
w
il
l
be
us
e
d
to
e
nha
nc
e
I
ndone
s
ia
’
s
ta
x
r
is
k
a
s
s
e
s
s
m
e
nt
m
e
c
ha
ni
s
m
.
2.
M
E
T
H
O
D
2.1. Dat
as
e
t
W
e
us
e
d
a
ta
f
r
om
61
r
is
k
e
le
m
e
nt
s
a
nd
e
xpl
a
na
to
r
y
va
r
ia
bl
e
s
(
f
e
a
tu
r
e
s
)
in
th
r
e
e
m
a
in
c
a
t
e
gor
ie
s
a
nd
f
iv
e
ty
pe
s
of
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
r
e
s
pons
e
va
r
ia
bl
e
s
c
a
r
r
ie
d
out
by
ta
xpa
ye
r
e
nt
it
ie
s
.
T
he
da
ta
ty
pe
is
a
num
e
r
ic
s
c
or
e
(
0
-
100)
w
it
h
a
to
ta
l
num
be
r
of
obs
e
r
va
ti
ons
of
498
e
nt
it
ie
s
,
a
s
e
xpl
a
in
e
d
in
T
a
bl
e
1.
A
c
c
or
di
ng
to
[
13]
,
th
e
r
e
s
a
m
pl
in
g
-
boot
s
tr
a
p
m
e
th
od
c
a
n
pr
ov
id
e
a
c
c
ur
a
te
m
ode
l
pe
r
f
or
m
a
nc
e
r
e
s
ul
ts
.
T
he
a
lg
or
it
hm
us
e
d
is
nonp
a
r
a
m
e
tr
ic
boot
s
tr
a
p,
w
hi
c
h
is
th
e
m
os
t
s
tr
a
ig
ht
f
or
w
a
r
d
a
nd
w
id
e
ly
us
e
d
m
e
th
od
[
14]
.
T
he
w
e
ig
ht
of
ta
x
a
voi
da
nc
e
or
ta
x
e
va
s
io
n
on
th
e
r
e
s
pons
e
va
r
i
a
bl
e
la
be
ls
in
th
is
s
tu
dy
is
c
ons
id
e
r
e
d
e
qua
l,
s
o
th
e
r
e
i
s
no pa
r
ti
c
ul
a
r
or
de
r
r
e
ga
r
di
ng pr
io
r
i
ty
l
a
be
ls
f
or
t
he
t
ype
s
of
t
hos
e
l
a
be
ls
.
T
a
bl
e
1.
E
xpl
a
na
to
r
y va
r
ia
bl
e
s
(
f
e
a
tu
r
e
s
)
a
nd t
a
r
ge
t
va
r
ia
bl
e
s
F
e
a
t
ur
e
T
a
r
ge
t
V
a
r
i
a
bl
e
s
1.
R
e
gi
s
t
r
a
t
i
on
di
m
e
ns
i
on
(
R
)
:
a
s
m
a
ny
a
s
22
e
l
e
m
e
nt
s
r
e
l
a
t
e
d t
o
:
a.
T
he
t
ype
s
of
bus
i
ne
s
s
e
nt
i
t
i
e
s
b.
T
a
x
obl
i
ga
t
i
ons
c.
A
c
c
ount
i
ng
s
ys
t
e
m
d.
P
r
oduc
t
i
on
c
om
pone
nt
s
2.
O
pe
r
a
t
i
ona
l
di
m
e
ns
i
on
(
O
)
:
a
s
m
a
ny
a
s
10
e
l
e
m
e
nt
s
r
e
l
a
t
e
d t
o
:
a.
C
ur
r
e
nt
a
s
s
e
t
r
a
t
i
o
b.
P
r
oduc
t
i
on
r
a
t
i
o
c.
A
udi
t
hi
s
t
or
y
3.
L
oc
a
l
di
m
e
ns
i
on
(
L
)
:
a
s
m
a
ny
a
s
29
e
l
e
m
e
nt
s
r
e
l
a
t
e
d
to
:
a.
F
a
c
t
or
y
s
e
c
ur
i
t
y s
ys
t
e
m
b.
S
a
l
e
s
t
r
a
ns
a
c
t
i
on s
y
s
t
e
m
c.
D
i
s
t
r
i
but
or
l
i
s
t
1.
L
a
be
l
1:
F
a
ke
,
s
e
c
ondha
nd
, or
no
t
a
x s
t
a
m
p doc
um
e
nt
s
2.
L
a
be
l
2:
W
r
ong
pe
r
s
ona
l
i
z
a
t
i
on
of
t
a
x s
t
a
m
p doc
um
e
nt
s
3.
L
a
be
l
3:
W
r
ong
pur
pos
e
t
a
x
s
t
a
m
p doc
um
e
nt
s
4.
L
a
be
l
4:
A
dm
i
ni
s
t
r
a
t
i
ve
s
a
nc
t
i
ons
or
f
i
ne
s
5.
L
a
be
l
5:
S
us
pe
ns
i
on
of
l
i
c
e
n
s
i
ng
or
r
e
voc
a
t
i
on of
f
a
c
i
l
i
t
i
e
s
2.2. Re
s
e
ar
c
h
m
e
t
h
od
T
hi
s
s
tu
dy
i
s
di
vi
de
d
in
to
s
e
ve
r
a
l
s
ta
ge
s
:
da
ta
e
xpl
or
a
ti
on,
M
L
C
m
ode
li
ng,
m
ode
l
e
va
lu
a
ti
on,
m
ode
l
pe
r
f
or
m
a
nc
e
c
om
pa
r
is
on,
a
nd
th
e
id
e
nt
i
f
ic
a
ti
on
of
dom
in
a
nt
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
f
a
c
to
r
s
ba
s
e
d
on
f
e
a
tu
r
e
im
por
ta
nc
e
va
lu
e
.
T
he
r
e
s
e
a
r
c
h
f
r
a
m
e
w
or
k
is
de
pi
c
te
d
in
F
ig
ur
e
1.
T
hr
oughout
th
e
r
e
s
e
a
r
c
h
pr
oc
e
s
s
,
w
e
us
e
d
P
yt
hon
a
nd
R
s
of
twa
r
e
.
T
hi
s
de
c
i
s
io
n
w
a
s
m
a
d
e
ba
s
e
d
on
th
e
e
a
s
e
of
in
te
gr
a
ti
on
w
it
h
th
e
M
L
C
li
br
a
r
y
a
nd t
he
ne
e
d f
or
unde
r
s
ta
nda
bl
e
a
n
a
ly
s
is
out
put
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
pr
ov
in
g t
he
r
is
k
p
r
of
il
e
of
I
ndone
s
ia
n e
nt
e
r
pr
i
s
e
t
ax
pay
e
r
s
u
s
in
g m
ul
ti
la
be
l
…
(
T
e
guh P
r
as
e
ty
o)
4325
F
ig
ur
e
1. R
e
s
e
a
r
c
h f
r
a
m
e
w
or
k
2.3.
D
at
a e
xp
lo
r
at
io
n
W
e
e
x
a
m
in
e
d
th
e
da
t
a
to
ge
t
a
br
oa
d
pi
c
tu
r
e
be
c
a
u
s
e
M
L
C
d
a
t
a
di
f
f
e
r
s
f
r
om
B
C
C
a
nd
M
C
C
d
a
ta
:
a
s
in
gl
e
ob
s
e
r
v
a
ti
on
c
a
n
b
e
c
l
a
s
s
if
i
e
d
in
t
o
m
ul
ti
pl
e
l
a
b
e
l
c
l
a
s
s
e
s
[
15]
.
W
e
ob
s
e
r
v
e
d
c
ha
r
a
c
t
e
r
is
ti
c
s
s
uc
h
a
s
la
be
l
di
s
tr
ib
ut
io
n,
la
b
e
l
r
e
l
a
ti
on
s
hi
ps
,
a
n
d
la
be
l
c
la
s
s
im
b
a
la
n
c
e
[
16]
.
T
he
M
L
C
m
e
th
od
m
e
a
s
ur
e
s
th
e
la
be
l
di
s
tr
i
but
io
n
us
in
g
th
e
c
a
r
di
na
l
it
y
v
a
lu
e
,
w
h
ic
h
is
th
e
a
v
e
r
a
g
e
num
be
r
of
a
c
t
iv
e
l
a
be
l
s
or
pa
ir
s
of
la
b
e
ls
f
or
e
a
c
h
ob
s
e
r
v
e
d
va
lu
e
, a
n
d
la
be
l
d
e
n
s
it
y
,
w
hi
c
h i
s
t
h
e
de
ns
i
ty
v
a
lu
e
of
e
a
c
h
l
a
b
e
l
i
n a
l
ow
e
r
di
m
e
n
s
io
n, d
e
s
c
r
ib
e
d i
n (
1)
.
(
)
=
1
|
|
∑
|
|
;
|
|
=
1
(
)
=
1
|
|
∑
|
|
|
|
|
|
=
1
(
1)
w
he
r
e
D
r
e
pr
e
s
e
nt
s
t
he
numb
e
r
of
obs
e
r
va
ti
ons
or
i
ns
ta
nc
e
s
a
nd
L
r
e
pr
e
s
e
nt
s
t
he
numbe
r
of
l
a
be
ls
.
N
e
xt
,
w
e
lo
ok
f
or
c
la
s
s
im
ba
la
nc
e
s
in
th
e
da
ta
.
O
ur
r
e
s
ul
ts
s
how
a
n
im
ba
la
nc
e
in
th
e
num
be
r
o
f
obs
e
r
va
ti
ons
be
tw
e
e
n
la
be
l
c
la
s
s
e
s
.
W
e
u
s
e
th
e
im
ba
la
nc
e
r
a
ti
o
(
I
R
L
bl
)
,
M
e
a
nI
R
,
a
nd
c
oe
f
f
ic
ie
nt
of
va
r
ia
ti
on
va
lu
e
s
t
o de
te
r
m
in
e
t
he
m
a
gni
tu
de
of
t
he
c
la
s
s
i
m
ba
la
n
c
e
, a
s
d
e
s
c
r
ib
e
i
n (
2)
.
(
)
=
(
′
)
(
∑
ℎ
(
′
,
|
|
=
1
)
∑
ℎ
(
,
)
|
|
=
1
,
ℎ
(
,
)
=
{
1
0
(
2)
=
1
|
|
∑
(
(
)
∈
)
C
la
s
s
i
m
ba
la
n
c
e
i
n M
L
C
i
s
c
la
s
s
if
ie
d a
s
i
m
ba
la
n
c
e
w
it
hi
n l
a
be
l
s
, i
m
ba
la
nc
e
be
twe
e
n m
ul
ti
pl
e
l
a
be
ls
,
a
nd
im
ba
la
nc
e
of
a
ll
la
b
e
ls
[
17]
.
I
t
c
a
n
be
a
ppr
oa
c
h
e
d
in
th
r
e
e
w
a
ys
:
u
s
in
g
da
ta
,
a
lg
or
it
hm
s
,
a
nd
a
c
om
bi
ne
d
a
ppr
oa
c
h.
A
c
c
or
di
ng
to
K
ur
ni
a
w
a
ti
a
nd
P
r
a
bow
o
[
18
]
,
bot
h
o
ve
r
s
a
m
pl
in
g
a
nd
unde
r
s
a
m
pl
in
g
m
e
th
ods
ha
ve
s
hor
tc
om
in
gs
,
s
uc
h
a
s
ove
r
f
it
ti
ng
th
e
m
ode
l
a
nd
lo
s
in
g
in
f
or
m
a
ti
on
in
obs
e
r
va
ti
ons
.
B
a
s
e
d
on
th
i
s
,
in
th
is
M
L
C
s
tu
dy
,
im
ba
la
nc
e
s
a
r
e
a
ddr
e
s
s
e
d
by s
tr
a
ti
f
yi
ng
th
e
tr
a
in
a
nd
te
s
t
da
ta
(
m
ul
ti
la
be
l
s
tr
a
ti
f
ie
d
s
huf
f
le
s
pl
it
)
.
I
f
th
e
da
ta
i
s
r
a
ndoml
y s
pl
it
, t
h
e
t
r
a
in
a
nd
t
e
s
t
da
ta
m
a
y h
a
ve
di
f
f
e
r
e
nt
pr
opor
ti
ons
. T
hi
s
i
s
be
c
a
u
s
e
t
he
M
L
C
’
s
l
a
be
l
s
a
r
e
li
nke
d
a
nd
ha
ve
a
r
e
l
a
ti
ons
hi
p
[
19]
.
T
h
e
s
tr
a
ti
f
ic
a
ti
on
a
ppr
oa
c
h
ha
s
be
e
n
s
how
n
to
im
pr
ove
c
r
os
s
-
va
li
da
ti
on
a
c
c
ur
a
c
y, bi
a
s
v
a
lu
e
s
,
a
nd c
la
s
s
if
ic
a
ti
on mode
ls
[
20]
.
M
L
C
e
xhi
bi
ts
c
ha
r
a
c
te
r
is
ti
c
s
of
one
or
m
or
e
c
or
r
e
la
te
d
l
a
be
l
c
la
s
s
e
s
.
T
he
da
ta
w
e
us
e
c
ont
a
in
s
s
e
ve
r
a
l
la
be
l
s
e
ts
.
W
e
m
e
a
s
ur
e
t
hi
s
us
in
g t
he
s
c
or
e
of
c
onc
ur
r
e
nc
e
a
m
o
ng i
m
ba
la
nc
e
d l
a
be
ls
v
a
lu
e
(
S
C
U
M
B
L
E
)
[
21]
,
e
xpl
a
in
e
d i
n (
3)
:
(
)
=
1
−
1
̅
̅
̅
̅
̅
̅
̅
̅
̅
̅
(
∏
|
|
=
1
)
(
1
|
|
)
(
3)
(
)
=
1
|
|
∑
(
)
|
|
=
1
2
.4.
M
u
lt
il
ab
e
l
m
e
t
h
od
s
T
he
m
ode
li
ng
s
ta
ge
us
e
s
th
e
s
c
ik
it
-
m
ul
ti
le
a
r
n
m
odul
e
in
P
yt
hon
s
of
twa
r
e
[
22]
.
T
he
M
L
C
m
e
th
od
u
s
e
s
th
r
e
e
a
ppr
oa
c
he
s
:
pr
obl
e
m
tr
a
ns
f
or
m
a
ti
on,
m
e
th
od
a
da
pt
a
ti
on,
a
nd
th
e
e
ns
e
m
bl
e
of
c
la
s
s
if
ie
r
s
[
23]
.
W
e
us
e
two me
th
ods
t
o c
om
p
a
r
e
e
a
c
h a
ppr
oa
c
h,
c
ons
id
e
r
in
g t
ha
t
e
a
c
h
a
ppr
oa
c
h ha
s
i
t
s
a
dva
nt
a
ge
s
a
nd dis
a
dva
nt
a
ge
s
.
T
he
f
ir
s
t
a
ppr
oa
c
h
is
th
e
pr
obl
e
m
tr
a
ns
f
or
m
a
ti
on,
w
hi
c
h
w
or
ks
by
c
onve
r
ti
ng
th
e
M
L
C
m
ode
l
in
to
a
s
im
pl
e
r
m
ode
l
(
B
C
C
/M
C
C
)
be
f
or
e
th
e
m
od
e
li
ng
s
ta
ge
.
I
n
th
is
a
ppr
oa
c
h,
w
e
us
e
th
e
bi
na
r
y
r
e
le
va
n
c
e
m
e
th
od
(
tr
a
ns
f
or
m
to
bi
na
r
y
c
la
s
s
if
ic
a
ti
on
pr
obl
e
m
)
in
tr
oduc
e
d
by
[
24
]
,
w
hi
c
h
w
or
ks
by
a
da
pt
in
g
th
e
one
-
ve
r
s
us
-
a
ll
a
ppr
oa
c
h
.
T
he
bi
na
r
y
r
e
l
e
va
nc
e
m
e
th
od
ha
s
s
e
ve
r
a
l
a
dva
nt
a
ge
s
,
in
c
lu
di
ng
a
s
im
pl
e
a
lg
or
it
hm
a
nd e
a
s
e
of
unde
r
s
t
a
ndi
ng
[
25]
.
W
e
a
ls
o
a
ppl
ie
d
th
e
la
be
l
pow
e
r
s
e
t
m
e
th
od
(
t
r
a
ns
f
or
m
to
m
ul
ti
c
la
s
s
c
la
s
s
if
ic
a
ti
on
pr
obl
e
m
)
in
tr
oduc
e
d
in
[
26]
.
T
he
w
or
ki
ng
p
r
in
c
ip
le
is
to
c
onve
r
t
th
e
M
L
C
m
ode
l
in
to
m
ul
ti
pl
e
M
C
C
m
ode
ls
.
E
a
c
h
pa
ir
of
obs
e
r
va
ti
ons
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4323
-
4333
4326
on
la
be
l
is
tr
e
a
te
d a
s
a
s
e
pa
r
a
te
obs
e
r
va
ti
on. Ac
c
or
di
ng
to
[
27]
,
if
th
e
r
e
is
a
s
tr
ong
c
or
r
e
la
ti
on
be
twe
e
n
la
be
l
s
,
th
e
l
a
be
l
pow
e
r
s
e
t
m
e
th
od i
s
r
e
c
om
m
e
nde
d.
T
he
s
e
c
ond
a
ppr
oa
c
h
is
a
lg
or
it
hm
a
da
pt
a
ti
on,
w
hi
c
h
w
or
ks
by
a
da
pt
in
g
a
c
la
s
s
if
ic
a
ti
on
m
ode
l
a
lg
or
it
hm
f
or
di
r
e
c
t
a
ppl
ic
a
ti
on
to
th
e
M
L
C
c
a
s
e
[
23]
.
T
hi
s
a
ppr
oa
c
h
r
e
li
e
s
on
a
r
ti
f
ic
ia
l
ne
ur
a
l
ne
twor
ks
,
s
pe
c
if
ic
a
ll
y
m
ul
ti
la
be
l
-
a
da
pt
iv
e
r
e
s
ona
nc
e
a
s
s
oc
ia
ti
ve
m
a
p
(
M
L
-
A
R
A
M
)
ba
s
e
d
on
de
e
p
le
a
r
ni
ng
m
e
th
od.
A
c
c
or
di
ng
to
[
28]
,
th
e
w
or
ki
ng
pr
in
c
ip
le
of
M
L
-
A
R
A
M
is
to
us
e
a
da
pt
iv
e
r
e
s
on
a
nc
e
th
e
or
y
(
A
R
T
)
to
s
pe
e
d
up
m
ode
l
c
la
s
s
if
ic
a
ti
on.
[
29]
e
xpl
a
in
s
th
a
t
th
e
M
L
-
A
R
A
M
m
e
th
od
e
m
pl
oys
f
uz
z
y
lo
gi
c
pr
in
c
ip
le
s
,
w
it
h
A
R
T
be
in
g a
f
uz
z
y m
od
e
l
a
r
ti
f
ic
ia
l
ne
ur
a
l
ne
twor
k m
odul
e
w
it
h t
w
o
hi
dde
n l
a
ye
r
s
, F
1
a
nd F
2. T
h
e
l
e
a
r
ni
ng pr
oc
e
s
s
in
t
he
A
R
T
a
a
nd A
R
T
b l
a
ye
r
s
i
s
d
e
s
c
r
ib
e
d a
s
(
4)
:
(
)
=
(
∧
(
)
)
+
(
1
−
)
(
)
(
4)
W
is
t
he
w
e
ig
ht
of
th
e
ne
ur
a
l
ne
t
w
or
k, A
is
th
e
i
nput
(
e
xpl
a
na
to
r
y
va
r
ia
bl
e
/f
e
a
tu
r
e
s
)
, a
nd
β
i
s
th
e
le
a
r
ni
ng
r
a
te
(
0,1)
.
W
e
a
ls
o
e
m
pl
oy
th
e
m
ul
ti
l
a
b
e
l
k
-
n
e
a
r
e
s
t
ne
i
ghbor
(
M
L
-
kN
N
)
m
e
th
od
pr
opo
s
e
d
by
[
30]
,
w
hi
c
h
i
s
th
e
m
o
s
t
e
f
f
e
c
ti
v
e
o
bs
e
r
va
ti
on
-
ba
s
e
d
M
L
C
m
e
th
od.
T
he
M
L
-
kN
N
a
lg
or
it
h
m
us
e
s
th
e
k
-
ne
a
r
e
s
t
n
e
ig
hbor
(
kN
N
)
a
lg
or
it
hm
in
c
l
us
t
e
r
a
na
l
ys
i
s
,
w
hi
c
h
i
s
a
no
npa
r
a
m
e
tr
i
c
m
e
th
od
t
h
a
t
us
e
s
t
h
e
c
lo
s
e
di
s
t
a
n
c
e
be
t
w
e
e
n o
bs
e
r
va
ti
ons
.
T
h
e
r
e
i
s
a
m
odi
f
i
c
a
ti
o
n of
t
h
e
M
L
-
kN
N
m
e
th
o
d t
h
a
t
u
s
e
s
t
h
e
B
a
ye
s
pr
ob
a
bi
li
ty
pr
in
c
ip
le
,
a
s
e
x
pl
a
in
e
d i
n (
5)
:
⃗
(
)
=
arg
∈
{
0
,
1
}
(
)
.
(
⃗
⃗
⃗
(
)
|
)
(
⃗
⃗
⃗
(
)
)
(
5)
W
he
r
e
t
is
a
ne
w
da
ta
s
e
t,
is
a
n
e
ve
nt
w
he
th
e
r
t
ha
s
a
la
be
l
or
not
,
a
nd
⃗
(
)
is
th
e
num
be
r
of
e
ve
nt
s
a
s
m
a
ny
a
s
K
ne
a
r
e
s
t
ne
ig
hbor
s
of
t
w
it
h a
l
a
be
l
l
.
T
he
th
ir
d
a
ppr
oa
c
h
is
th
e
e
ns
e
m
bl
e
of
c
la
s
s
if
ie
r
,
w
hi
c
h
ha
s
th
e
w
or
ki
ng
pr
in
c
ip
le
o
f
m
ode
li
ng
M
L
C
us
in
g
s
e
ve
r
a
l
ba
s
ic
pr
obl
e
m
tr
a
ns
f
or
m
a
ti
on
m
e
th
ods
a
nd
a
lg
or
it
hm
/a
da
pt
a
ti
on
m
e
th
ods
s
im
ul
ta
ne
ou
s
ly
,
s
uc
h
a
s
boos
ti
ng
or
ba
ggi
ng.
I
n
th
is
a
ppr
oa
c
h,
w
e
us
e
th
e
l
a
be
l
s
p
a
c
e
p
a
r
ti
ti
oni
ng
c
la
s
s
if
ie
r
m
e
th
od,
w
it
h
th
e
w
or
ki
ng
pr
in
c
ip
le
of
di
vi
di
ng
pa
ir
s
of
r
e
s
pons
e
va
r
ia
bl
e
obs
e
r
va
ti
on
s
in
to
s
e
ve
r
a
l
s
m
a
ll
c
lu
s
te
r
s
a
nd
bui
ld
in
g
a
m
ul
ti
c
la
s
s
c
la
s
s
if
ic
a
ti
on mode
l
us
in
g t
he
l
a
be
l
s
pa
c
e
c
lu
s
te
r
e
r
m
e
th
od s
e
pa
r
a
te
ly
f
or
e
a
c
h pa
ir
of
obs
e
r
va
ti
ons
us
in
g
a
hybr
id
pa
r
ti
ti
on
a
ppr
oa
c
h
[
31]
.
T
he
n,
w
e
a
ls
o
us
e
d
t
he
r
a
ndom
la
be
l
s
pa
c
e
p
a
r
ti
ti
oni
ng
w
it
h
la
be
l
pow
e
r
s
e
t
m
e
th
od
(
R
A
kE
L
d)
,
w
hi
c
h
w
or
ks
by
r
a
ndomi
z
in
g
s
e
ve
r
a
l
pa
ir
s
of
obs
e
r
va
ti
ons
f
r
om
la
be
ls
a
nd
bui
ld
in
g
a
m
ul
ti
c
la
s
s
c
la
s
s
if
ic
a
ti
on
m
ode
l
u
s
in
g
th
e
la
be
l
p
ow
e
r
s
e
t
m
e
th
od
s
e
pa
r
a
te
ly
f
or
e
a
c
h
p
a
ir
of
obs
e
r
va
ti
ons
[
23]
.
T
he
R
a
kE
L
d
m
e
th
od
ha
s
a
n
a
dva
nt
a
ge
ove
r
th
e
la
be
l
pow
e
r
s
e
t
m
e
th
od:
be
c
a
u
s
e
it
c
a
n
pr
e
di
c
t
la
be
l
pa
ir
s
th
a
t
h
a
ve
not
y
e
t
to
a
ppe
a
r
in
th
e
da
ta
.
I
t
f
a
c
il
it
a
te
s
t
he
ge
ne
r
a
li
z
a
ti
on
of
pr
e
di
c
ti
ons
w
he
n
th
e
da
ta
s
a
m
pl
e
doe
s
not
y
e
t
r
e
pr
e
s
e
nt
popula
ti
on c
ha
r
a
c
t
e
r
is
ti
c
s
[
32]
.
2.5
.
M
od
e
l
e
val
u
at
io
n
W
e
c
om
pa
r
e
d
th
e
m
ode
l
pe
r
f
or
m
a
nc
e
us
in
g
th
e
c
r
os
s
-
va
li
da
ti
on
m
e
th
od
by
pa
r
ti
ti
oni
ng
th
e
da
ta
in
to
tr
a
in
a
nd
te
s
t
d
a
ta
in
10
pa
r
ti
ti
ons
.
F
ur
th
e
r
m
or
e
,
c
r
os
s
-
va
li
da
ti
on
c
a
n
be
us
e
d
to
e
s
ti
m
a
te
th
e
m
ode
l
a
ve
r
a
g
e
pr
e
di
c
ti
on
e
r
r
or
w
he
n
th
e
tr
a
in
a
nd
te
s
t
da
t
a
a
r
e
r
a
ndoml
y
s
e
le
c
te
d
f
r
om
th
e
ov
e
r
a
ll
da
ta
[
33]
.
M
ode
l
pe
r
f
or
m
a
nc
e
e
va
lu
a
ti
on
yi
e
ld
s
th
e
be
s
t
m
ode
l
ba
s
e
d
on
th
e
F
1
-
s
c
or
e
a
nd
ha
m
m
in
g
lo
s
s
.
T
he
M
L
C
m
ode
l
pe
r
f
or
m
a
nc
e
is
e
va
lu
a
te
d
di
f
f
e
r
e
nt
ly
f
r
om
t
he
B
C
C
or
M
C
C
m
o
de
ls
due
to
th
e
di
f
f
e
r
e
nt
ty
pe
s
a
nd
s
tr
uc
tu
r
e
s
of
r
e
s
pons
e
v
a
r
ia
bl
e
s
[
34]
.
I
n
a
ddi
ti
on
to
m
e
th
od
s
e
le
c
ti
on,
m
ode
l
pe
r
f
or
m
a
nc
e
is
a
ls
o
in
f
lu
e
nc
e
d
by
da
ta
c
ha
r
a
c
te
r
is
ti
c
s
a
nd l
a
b
e
l
c
la
s
s
s
e
pa
r
a
ti
on pa
tt
e
r
ns
[
35]
. W
e
us
e
t
he
F
1
-
s
c
or
e
, a
la
be
l
-
ba
s
e
d m
e
tr
ic
, c
ons
id
e
r
in
g
th
e
pr
e
c
is
io
n a
nd r
e
c
a
ll
va
lu
e
s
a
s
(
6)
[
36]
:
1
−
=
2
×
×
+
;
or
:
1
−
=
1
∑
2
|
∩
̂
|
|
̂
|
+
|
|
=
1
(
6)
r
e
pr
e
s
e
nt
s
th
e
num
b
e
r
of
obs
e
r
va
ti
on
s
in
th
e
da
ta
,
is
th
e
tr
ue
/
a
c
t
ua
l
l
a
be
l
va
l
ue
,
a
n
d
̂
is
th
e
pr
e
di
c
te
d
va
l
ue
.
T
he
in
te
r
s
e
c
ti
on
of
a
nd
̂
in
di
c
a
te
s
th
e
n
um
be
r
of
la
b
e
l
s
pr
e
d
ic
t
e
d
c
or
r
e
c
tl
y
by
t
he
m
ode
l.
T
he
hi
ghe
r
th
e
F1
-
s
c
or
e
va
lu
e
,
th
e
be
tt
e
r
th
e
m
od
e
l
pe
r
f
or
m
a
nc
e
[
37]
.
T
he
F1
-
s
c
or
e
,
a
c
c
or
di
ng
to
[
38]
w
hi
c
h
c
o
ns
id
e
r
s
pr
e
c
i
s
io
n
a
nd
r
e
c
a
ll
a
s
m
e
a
s
ur
e
s
of
m
o
de
l
goodn
e
s
s
,
i
s
r
obu
s
t
to
in
di
c
a
ti
on
s
of
i
m
ba
l
a
nc
e
d l
a
be
ls
in
t
h
e
d
a
ta
.
I
n
a
ddi
ti
o
n
to
th
e
F
1
-
s
c
or
e
,
w
e
u
s
e
th
e
h
a
m
m
in
g
lo
s
s
,
a
n
I
n
s
ta
n
c
e
-
ba
s
e
d
m
e
tr
i
c
,
by
e
v
a
lu
a
ti
ng
th
e
num
b
e
r
of
c
la
s
s
if
ic
a
ti
on e
r
r
or
s
of
t
he
l
a
be
l/
c
la
s
s
r
e
s
pon
s
e
va
r
ia
bl
e
pa
ir
s
f
r
om
t
he
r
e
s
ul
ti
ng mode
l,
a
s
f
or
m
ul
a
te
d i
n (
7)
:
=
1
∑
|
≠
̂
|
=
1
(
7)
r
e
pr
e
s
e
nt
s
th
e
m
a
xi
m
um
num
be
r
of
la
be
ls
on
th
e
r
e
s
pon
s
e
va
r
ia
bl
e
.
T
h
e
f
e
w
e
r
m
is
pr
e
di
c
t
la
b
e
ls
,
th
e
lo
w
e
r
th
e
ha
m
m
in
g l
os
s
, a
nd t
he
hi
ghe
r
t
he
m
ode
l
pe
r
f
or
m
a
nc
e
[
39]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
pr
ov
in
g t
he
r
is
k
p
r
of
il
e
of
I
ndone
s
ia
n e
nt
e
r
pr
i
s
e
t
ax
pay
e
r
s
u
s
in
g m
ul
ti
la
be
l
…
(
T
e
guh P
r
as
e
ty
o)
4327
2.6
.
M
od
e
l
p
e
r
f
or
m
an
c
e
c
o
m
p
ar
is
on
We
c
om
pa
r
e
th
e
m
ode
l
pe
r
f
or
m
a
nc
e
us
in
g
th
e
a
na
ly
s
i
s
of
va
r
ia
nc
e
m
e
th
od.
T
he
M
L
C
m
e
th
od
a
nd
da
ta
s
e
pa
r
a
ti
on
m
e
th
od
i
s
c
om
pa
r
e
d
by
th
e
m
e
a
n
va
lu
e
of
th
e
M
L
C
m
ode
l
pe
r
f
or
m
a
nc
e
(
F
1
-
s
c
or
e
a
nd
ha
m
m
in
g
lo
s
s
)
.
T
he
c
om
pa
r
is
on
e
ns
ur
e
s
th
a
t
th
e
di
ve
r
s
it
y
be
tw
e
e
n
gr
oups
is
he
te
r
oge
ne
ous
u
s
in
g
th
e
a
na
ly
s
i
s
of
va
r
ia
nc
e
m
e
th
od
[
40]
, a
s
f
or
m
ul
a
te
d i
n (
8)
:
=
+
+
+
(
)
+
(
8)
de
not
e
t
he
M
L
C
m
e
th
ods
a
nd
a
r
e
t
he
da
ta
s
e
pa
r
a
ti
on me
th
ods
.
2.7.
F
e
at
u
r
e
i
m
p
or
t
an
c
e
I
n
th
e
f
ol
lo
w
in
g
s
ta
ge
,
w
e
id
e
nt
if
y
r
is
k e
le
m
e
nt
f
a
c
to
r
s
th
a
t
e
nc
our
a
ge
a
ta
xpa
y
e
r
e
nt
it
y
to
e
ng
a
ge
in
ta
x a
voi
da
nc
e
or
t
a
x e
va
s
io
n us
in
g t
he
f
e
a
tu
r
e
i
m
por
ta
nc
e
va
lu
e
.
[
41]
e
xpl
a
in
e
d t
ha
t
m
a
c
hi
ne
l
e
a
r
ni
ng mode
ls
m
us
t
be
in
te
r
pr
e
te
d,
e
ve
n
th
ough
s
om
e
m
ode
ls
ha
ve
va
r
yi
ng
de
gr
e
e
s
of
di
f
f
ic
ul
ty
.
S
om
e
m
e
th
ods
,
s
uc
h
a
s
tr
e
e
-
ba
s
e
d m
e
th
ods
, a
r
e
s
tr
a
ig
ht
f
or
w
a
r
d t
o unde
r
s
ta
nd
. T
hi
s
s
tu
dy e
m
pl
oys
t
he
pe
r
m
ut
a
ti
on
f
e
a
t
ur
e
i
m
por
ta
n
c
e
va
lu
e
,
w
hi
c
h
d
e
s
c
r
ib
e
s
t
he
e
f
f
e
c
t
of
e
a
c
h
e
xpl
a
na
to
r
y
v
a
r
ia
bl
e
(
f
e
a
tu
r
e
s
)
on
th
e
pr
e
di
c
ti
on
r
e
s
ul
t
s
.
T
h
e
a
dv
a
nt
a
g
e
of
th
i
s
a
ppr
oa
c
h
is
th
a
t
th
e
in
t
e
r
pr
e
t
a
ti
on
i
s
s
im
pl
e
to
und
e
r
s
t
a
nd
[
42]
.
A
c
c
r
odi
ng
t
o
[
43]
,
th
e
pe
r
m
ut
a
ti
on
m
e
th
od
is
w
id
e
ly
us
e
d
to
de
te
r
m
in
e
th
e
f
e
a
tu
r
e
im
por
ta
nc
e
of
a
m
a
c
hi
ne
le
a
r
ni
ng
m
ode
l.
M
a
c
hi
ne
le
a
r
ni
ng
m
ode
ls
th
a
t
us
e
hi
gh
-
di
m
e
ns
io
na
l
da
ta
be
c
om
e
m
or
e
c
om
pl
e
x.
S
im
pl
e
m
ode
ls
a
r
e
e
a
s
ie
r
to
in
te
r
pr
e
t
in
de
c
is
io
n
-
m
a
ki
ng,
e
s
pe
c
ia
ll
y
w
h
e
n
c
on
s
id
e
r
in
g
good
m
ode
l
p
e
r
f
or
m
a
nc
e
.
D
im
e
n
s
io
na
li
ty
r
e
duc
ti
on
is
a
m
e
th
od
f
or
c
onve
r
ti
ng
hi
gh
-
di
m
e
ns
io
na
l
da
ta
in
to
lo
w
e
r
-
di
m
e
ns
io
na
l
da
ta
,
th
a
t
e
m
pl
oy
s
two
a
ppr
oa
c
he
s
:
f
e
a
tu
r
e
s
e
l
e
c
ti
on
a
nd
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
[
44]
.
N
e
xt
,
w
e
us
e
th
e
s
e
dom
in
a
nt
f
a
c
to
r
s
to
r
e
duc
e
r
is
k
e
le
m
e
nt
s
,
m
a
ki
ng
th
e
m
e
a
s
ie
r
to
in
te
r
pr
e
t
a
nd de
c
id
e
.
3.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
O
ur
f
in
di
ngs
a
r
e
or
ga
ni
z
e
d
in
to
th
r
e
e
:
da
ta
e
xpl
or
a
ti
on,
m
ode
l
e
va
lu
a
ti
on,
a
nd
pe
r
f
or
m
a
nc
e
c
om
pa
r
is
on,
a
s
w
e
ll
a
s
f
e
a
tu
r
e
im
por
ta
nc
e
a
nd
s
e
le
c
ti
on.
T
hi
s
s
tu
dy
a
im
s
to
a
ddr
e
s
s
th
e
s
tu
dy
obj
e
c
ti
ve
s
,
w
hi
c
h
in
c
lu
de
de
a
li
ng
w
it
h
c
la
s
s
im
ba
la
nc
e
,
c
om
pa
r
in
g
M
L
C
m
e
th
ods
,
a
nd
id
e
nt
if
yi
ng
dom
in
a
nt
f
a
c
to
r
s
th
a
t
e
nc
our
a
ge
a
ta
xpa
ye
r
e
nt
it
y
to
e
nga
ge
in
ta
x
a
voi
da
nc
e
or
ta
x
e
va
s
io
n.
W
e
pr
ovi
de
a
di
s
c
us
s
io
n
of
e
a
c
h
a
na
ly
s
is
r
e
s
ul
t
to
s
uppor
t
th
e
s
tu
dy
f
in
di
ngs
.
3.1. Dat
a e
xp
lo
r
at
io
n
A
m
ong
th
e
f
iv
e
c
la
s
s
e
s
of
da
ta
la
be
ls
,
la
be
l
4
dom
in
a
te
s
w
it
h
a
f
r
e
que
nc
y
of
m
or
e
th
a
n
200
obs
e
r
va
ti
ons
,
w
hi
le
la
be
l
1
ha
s
th
e
f
e
w
e
s
t.
M
e
a
nw
hi
le
,
th
e
la
be
l
pa
ir
s
w
it
h
th
e
hi
ghe
s
t
f
r
e
que
nc
y
of
obs
e
r
va
ti
ons
a
r
e
la
be
l
4
a
nd
2,
a
s
w
e
ll
a
s
th
e
pa
ir
of
la
be
ls
2
a
nd
4.
L
a
be
ls
w
it
h
m
or
e
pa
ir
s
te
nd
to
ha
ve
a
lo
w
e
r
f
r
e
que
nc
y
of
obs
e
r
va
ti
ons
.
T
he
e
m
pi
r
ic
a
l
da
t
a
c
ont
a
in
s
21
p
a
ir
s
of
la
be
ls
(
s
in
gl
e
a
nd
pa
ir
e
d
la
be
l
s
)
.
A
s
th
e
num
be
r
of
la
be
l
pa
ir
s
,
th
e
M
L
C
m
ode
li
ng
s
ta
ge
be
c
om
e
m
or
e
c
om
pl
ic
a
te
d.
F
ig
ur
e
2
de
pi
c
ts
th
e
la
be
l
di
s
tr
ib
ut
io
n
in
e
m
pi
r
ic
a
l
da
ta
.
A
da
ta
c
a
r
di
na
li
ty
va
lu
e
of
1.04
in
di
c
a
te
s
th
a
t
one
obs
e
r
va
ti
on
r
e
s
ul
t
s
in
a
n
a
ve
r
a
ge
of
one
to
two
la
be
ls
.
A
da
ta
de
ns
it
y
va
lu
e
of
0.21
in
d
ic
a
te
s
th
a
t
on
a
s
c
a
le
of
1
–
10,
21%
o
f
la
be
ls
ha
ve
th
e
gr
e
a
te
s
t
num
be
r
of
ob
s
e
r
va
ti
ons
.
T
he
hi
gh
e
r
th
e
c
a
r
di
na
li
ty
a
nd
de
ns
it
y
va
lu
e
s
,
th
e
m
or
e
la
be
l
c
a
te
gor
ie
s
a
r
e
ge
ne
r
a
te
d f
r
om
a
s
in
gl
e
da
ta
obs
e
r
va
ti
on, whic
h c
om
pl
ic
a
te
s
t
he
M
L
C
m
ode
li
ng s
ta
g
e
.
F
ig
ur
e
2. F
r
e
que
nc
y of
e
a
c
h l
a
be
l
a
nd l
a
be
l
pa
ir
s
f
or
e
m
pi
r
ic
a
l
da
ta
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4323
-
4333
4328
F
ig
ur
e
3 i
ll
us
tr
a
te
s
c
la
s
s
i
m
ba
l
a
nc
e
c
om
pa
r
in
g t
he
numbe
r
of
o
bs
e
r
va
ti
ons
be
tw
e
e
n l
a
be
l
c
la
s
s
e
s
. W
e
c
a
n
s
e
e
vi
s
ua
l
s
ig
ns
of
c
l
a
s
s
im
ba
la
nc
e
on
f
our
of
th
e
f
iv
e
la
be
ls
in
th
e
da
ta
.
C
a
te
gor
y
0
(
doe
s
not
e
nga
g
e
in
ta
x
a
voi
da
nc
e
or
ta
x
e
va
s
io
n)
dom
in
a
te
s
a
ll
la
be
ls
.
L
a
be
l
1
ha
s
t
he
hi
ghe
s
t
I
R
L
bl
va
lu
e
a
t
8.40.
M
e
a
n
w
hi
le
,
th
e
M
e
a
nI
R
da
ta
v
a
lu
e
i
s
3.49, me
a
ni
ng t
he
a
ve
r
a
ge
I
R
L
bl
a
c
r
os
s
a
ll
l
a
be
ls
i
s
one
ve
r
s
u
s
3.49.
F
ig
ur
e
3. C
om
pa
r
is
on of
t
he
numbe
r
of
c
la
s
s
obs
e
r
va
ti
ons
b
e
tw
e
e
n l
a
be
ls
We
di
s
c
ove
r
e
d
th
a
t
th
e
la
be
l
c
la
s
s
w
it
h
th
e
hi
ghe
s
t
r
e
la
ti
ons
hi
p
va
lu
e
w
a
s
la
be
l
2,
f
ol
lo
w
e
d
by
la
be
l
4
w
it
h S
C
U
M
B
L
E
(
i)
va
lu
e
s
of
1.82 a
nd 1.35. W
e
c
a
lc
ul
a
te
t
he
a
ve
r
a
ge
va
lu
e
of
t
he
r
e
la
ti
ons
hi
p be
twe
e
n l
a
be
l
s
w
it
h
th
e
S
C
U
M
B
L
E
(
D
)
va
lu
e
,
w
he
r
e
th
e
hi
ghe
r
th
e
S
C
U
M
B
L
E
(
D
)
va
lu
e
,
th
e
m
or
e
c
om
pl
ic
a
te
d
th
e
M
L
C
m
ode
li
ng.
T
he
S
C
U
M
B
L
E
(
D
)
va
lu
e
in
th
e
da
ta
is
0.03,
w
hi
c
h
is
not
pa
r
ti
c
ul
a
r
ly
hi
gh
c
om
pa
r
e
d
to
ot
he
r
m
ul
ti
la
be
l
da
ta
ty
pe
s
,
s
uc
h
a
s
im
a
g
e
s
a
nd
te
xt
da
ta
.
T
h
e
hi
ghe
r
th
e
S
C
U
M
B
L
E
va
lu
e
,
th
e
m
or
e
obs
e
r
va
ti
ons
pr
o
duc
e
t
he
l
a
be
l
s
im
ul
ta
ne
ou
s
ly
. F
ig
ur
e
4 de
pi
c
ts
t
he
r
e
la
ti
ons
hi
p be
twe
e
n l
a
be
l
s
a
s
s
e
e
n
t
hr
ough the
numbe
r
of
obs
e
r
va
ti
ons
, w
it
h l
a
be
l
4 l
e
a
di
ng t
he
l
a
be
l
w
it
h t
he
m
o
s
t
obs
e
r
va
ti
ons
.
F
ig
ur
e
4. R
e
la
ti
ons
hi
ps
be
twe
e
n l
a
be
ls
W
e
a
ddr
e
s
s
c
la
s
s
im
ba
la
n
c
e
in
th
e
d
a
ta
u
s
in
g
a
s
tr
a
ti
f
ie
d
da
t
a
s
e
pa
r
a
ti
on
m
e
th
od
(
m
ul
ti
a
be
l
s
tr
a
ti
f
ie
d
s
huf
f
le
s
pl
it
)
to
pr
oduc
e
ba
la
nc
e
d
c
la
s
s
s
tr
a
ta
bot
h
tr
a
in
a
nd
te
s
t
da
ta
.
F
ur
th
e
r
m
o
r
e
,
th
e
la
be
l
c
la
s
s
s
tr
a
ta
ge
ne
r
a
te
d
in
th
e
tr
a
in
a
nd
te
s
t
da
ta
a
r
e
s
im
il
a
r
to
th
os
e
in
th
e
i
ni
ti
a
l
da
ta
,
a
s
s
how
n
in
F
ig
ur
e
s
5
(
a
)
a
nd
5(
b)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
pr
ov
in
g t
he
r
is
k
p
r
of
il
e
of
I
ndone
s
ia
n e
nt
e
r
pr
i
s
e
t
ax
pay
e
r
s
u
s
in
g m
ul
ti
la
be
l
…
(
T
e
guh P
r
as
e
ty
o)
4329
r
e
s
pe
c
ti
ve
ly
.
T
he
b
a
la
nc
e
of
th
e
s
e
s
tr
a
ta
a
ll
ow
s
th
e
M
L
C
m
ode
l
to
pr
oduc
e
good
r
e
s
ul
ts
w
hi
le
be
in
g
un
a
f
f
e
c
te
d
by
c
la
s
s
im
ba
la
nc
e
.
A
c
c
or
di
ng
to
[
45]
,
if
th
e
S
C
U
M
B
L
E
(
D
)
va
lu
e
is
<
0.1,
r
e
s
a
m
pl
in
g
(
unde
r
s
a
m
pl
in
g/
ove
r
s
a
m
pl
in
g)
i
s
unne
c
e
s
s
a
r
y t
o a
ddr
e
s
s
c
la
s
s
i
m
ba
la
nc
e
.
(
a
)
(
b)
F
ig
ur
e
5.
C
la
s
s
s
tr
a
ta
b
e
twe
e
n l
a
be
l
s
on
(
a
)
i
ni
ti
a
l
da
ta
a
nd (
b)
s
tr
a
ti
f
ic
a
ti
on r
e
s
ul
ts
3.2. M
od
e
l
p
e
r
f
or
m
an
c
e
e
val
u
at
io
n
W
it
h
a
100
-
r
e
pe
ti
ti
on
r
e
s
a
m
pl
in
g
s
c
e
na
r
io
s
how
n
in
F
ig
ur
e
6
.
T
he
F1
-
s
c
or
e
of
m
ode
l
pe
r
f
or
m
a
nc
e
is
s
how
n
in
F
ig
ur
e
6(
a
)
w
h
e
r
e
a
s
th
e
h
a
m
m
in
g
is
s
how
n
in
F
ig
ur
e
6(
b)
,
w
e
di
s
c
ove
r
e
d
th
a
t
th
e
M
L
-
A
R
A
M
m
e
th
od
ha
s
e
xc
e
ll
e
nt
m
ode
l
pe
r
f
or
m
a
nc
e
,
a
s
e
vi
de
nc
e
d
by
a
hi
ghe
r
pe
r
f
or
m
a
nc
e
m
e
a
n
a
nd
le
s
s
va
r
ia
bi
li
ty
w
he
n
c
om
pa
r
e
d
to
ot
he
r
m
e
th
ods
us
in
g
th
e
F
1
-
s
c
or
e
a
nd
ha
m
m
in
g
lo
s
s
.
T
h
e
F
1
-
s
c
or
e
obt
a
in
e
d
u
s
in
g
M
L
-
A
R
A
M
m
e
th
od
is
0.9,
i
ndi
c
a
ti
ng
th
a
t
th
e
M
L
-
A
R
A
M
m
e
th
od
pe
r
f
or
m
s
a
lm
os
t
pe
r
f
e
c
tl
y
M
L
C
c
la
s
s
if
ic
a
ti
on
pe
r
f
or
m
a
nc
e
.
M
e
a
nw
hi
le
,
th
e
M
L
-
A
R
A
M
m
e
th
od
pr
oduc
e
s
a
m
in
im
a
l
th
e
ha
m
m
in
g
lo
s
s
,
le
s
s
th
a
n
0.1.
I
n
a
ddi
ti
on
to
M
L
-
A
R
A
M
m
e
th
od,
th
e
la
be
l
pow
e
r
s
e
t
m
e
th
od
w
o
r
ks
w
e
ll
w
it
h
a
s
tr
a
ti
f
ie
d
da
ta
s
e
pa
r
a
ti
on.
T
he
s
tr
a
ti
f
ie
d
da
ta
s
e
p
a
r
a
ti
on
m
e
th
od
ha
s
b
e
e
n
s
how
n
to
im
pr
ove
m
ode
l
pe
r
f
or
m
a
nc
e
va
lu
e
s
in
bot
h
c
a
s
e
tr
a
ns
f
or
m
a
ti
on
a
nd
m
e
th
od
a
da
pt
a
ti
on
a
ppr
oa
c
h
e
s
.
H
ow
e
ve
r
,
t
he
m
ode
l
pe
r
f
or
m
a
nc
e
va
lu
e
dr
ops
s
li
ght
ly
in
th
e
e
ns
e
m
bl
e
of
c
la
s
s
if
ie
r
a
ppr
oa
c
h.
T
hi
s
is
m
os
t
li
ke
ly
due
t
o
th
e
us
e
of
m
ul
ti
pl
e
m
e
th
ods
a
t
onc
e
in
th
e
e
ns
e
m
bl
e
of
c
la
s
s
if
ie
r
a
ppr
oa
c
h,
w
hi
c
h
c
a
n
ha
ndl
e
c
la
s
s
im
ba
l
a
nc
e
w
it
hout
f
ir
s
t
s
tr
a
ti
f
yi
ng
th
e
tr
a
in
a
nd
te
s
t
da
ta
.
W
it
h
th
e
s
e
f
in
di
ngs
,
w
e
pr
opos
e
th
a
t
in
f
ut
ur
e
s
im
il
a
r
s
tu
dy
,
r
e
s
e
a
r
c
he
r
s
us
e
a
s
tr
a
ti
f
ie
d
da
ta
s
e
pa
r
a
ti
on
m
e
th
od t
o r
e
duc
e
t
he
i
m
pa
c
t
of
c
la
s
s
i
m
ba
la
nc
e
i
n t
he
M
L
C
m
e
th
od on the
r
e
s
ul
ts
of
[
20]
.
(
a
)
(
b)
F
ig
ur
e
6. C
om
pa
r
is
on of
m
ode
l
pe
r
f
or
m
a
nc
e
ba
s
e
d on (
a
)
F
1
-
s
c
or
e
a
nd (
b)
ha
m
m
in
g l
os
s
3.3. M
od
e
l
p
e
r
f
or
m
an
c
e
c
o
m
p
ar
is
on
T
a
bl
e
2
i
s
a
n
a
ly
s
is
of
va
r
i
a
nc
e
s
h
ow
s
th
e
m
ode
l
p
e
r
f
or
m
a
n
c
e
c
o
m
pa
r
is
on
r
e
s
ul
t.
B
a
s
e
d
on
t
he
p
-
va
lu
e
(
α=
0.0
5)
,
w
e
c
onc
l
ude
th
a
t
th
e
f
a
c
to
r
s
of
th
e
M
L
C
m
e
th
od,
d
a
ta
s
e
pa
r
a
ti
on
m
e
th
o
d,
a
nd
t
he
ir
i
nt
e
r
a
c
ti
on
be
twe
e
n
th
e
M
L
C
m
e
th
o
d
a
nd
th
e
d
a
ta
s
e
pa
r
a
ti
on
m
e
t
hod
a
f
f
e
c
t
th
e
m
e
a
n
v
a
lu
e
of
m
o
de
l
pe
r
f
or
m
a
nc
e
.
M
e
a
nw
hi
le
,
if
w
e
l
ook a
t
th
e
i
nt
e
r
a
c
ti
on plo
t
pr
e
s
e
nt
e
d i
n F
ig
u
r
e
7, i
t
i
s
c
onc
l
ude
d t
ha
t
th
e
e
f
f
e
c
t
of
da
ta
s
e
pa
r
a
ti
on me
th
ods
is
s
pe
c
if
ic
to
e
a
c
h M
L
C
m
e
th
od.
F
ig
ur
e
7(
a
)
s
how
s
th
e
in
te
r
a
c
t
io
n
ba
s
e
d
on
F
1
-
s
c
or
e
,
m
e
a
nw
hi
le
F
ig
ur
e
7(
b)
s
how
s
th
e
in
te
r
a
c
ti
on
ba
s
e
d
on
ha
m
m
in
g
lo
s
s
.
T
he
s
tr
a
ti
f
ie
d
da
ta
s
e
pa
r
a
ti
on
m
e
th
od
c
a
n
im
pr
ove
m
ode
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4323
-
4333
4330
pe
r
f
or
m
a
nc
e
in
m
os
t
M
L
C
m
e
th
ods
.
W
h
e
n
in
te
r
a
c
ti
ng
w
it
h
th
e
da
ta
s
e
pa
r
a
ti
on
m
e
th
od,
M
L
-
A
R
A
M
c
om
bi
ne
d
w
it
h
th
e
s
tr
a
ti
f
ie
d
da
ta
s
e
pa
r
a
ti
on
m
e
th
od
a
c
hi
e
ve
s
th
e
b
e
s
t
pe
r
f
or
m
a
nc
e
(
F
1
-
s
c
or
e
va
lu
e
of
0.955)
.
T
he
la
be
l
pow
e
r
s
e
t
c
om
bi
ne
d
w
it
h
th
e
s
tr
a
ti
f
ie
d
da
ta
s
e
p
a
r
a
ti
on
m
e
th
od
yi
e
ld
s
th
e
be
s
t
m
ode
l
pe
r
f
or
m
a
nc
e
ba
s
e
d
on
a
h
a
m
m
in
g
lo
s
s
of
0.063.
T
he
M
L
-
A
R
A
M
m
e
th
od,
w
h
e
n
c
om
bi
ne
d
w
it
h
s
tr
a
ti
f
ie
d
or
r
a
ndom
da
ta
s
e
pa
r
a
ti
on
m
e
th
ods
, s
ti
ll
pe
r
f
or
m
s
w
e
ll
i
n t
e
r
m
s
of
ha
m
m
in
g l
os
s
.
T
a
bl
e
2. M
od
e
l
pe
r
f
or
m
a
nc
e
a
na
ly
s
i
s
of
va
r
ia
nc
e
S
our
c
e
F1
-
s
c
or
e
H
a
m
m
i
ng l
os
s
F
-
s
t
a
t
i
s
t
i
c
s
P
-
va
l
ue
F
-
s
t
a
t
i
s
t
i
c
s
P
-
va
l
ue
M
L
C
m
e
t
hod
s
12158.84
0.000*
6514.75
0
.
000*
D
a
t
a
s
e
pa
r
a
t
i
on m
e
t
hods
4589.42
0.000*
4213.44
0
.
000*
M
L
C
m
e
t
hod
s
12158.84
0.000*
6514.75
0
.
000*
(
a
)
(
b
)
F
ig
ur
e
7. I
nt
e
r
a
c
ti
on M
L
C
a
nd da
ta
s
e
pa
r
a
ti
on me
th
ods
b
a
s
e
d o
n
(
a
)
F1
-
s
c
or
e
a
nd
(
b)
h
a
m
m
in
g
l
os
s
U
s
in
g
a
pos
t
hoc
te
s
t
on
th
e
M
L
C
m
e
th
od
f
a
c
to
r
s
th
a
t
ha
ve
be
e
n
s
e
pa
r
a
te
d
ba
s
e
d
on
da
ta
s
e
pa
r
a
ti
on
m
e
th
ods
pr
e
s
e
nt
e
d
in
T
a
bl
e
3
,
w
e
c
on
c
lu
de
th
a
t
M
L
-
A
R
A
M
ha
s
th
e
be
s
t
pe
r
f
or
m
a
nc
e
in
th
e
th
r
e
e
c
ondi
ti
ons
of
c
om
bi
na
ti
ons
.
M
L
-
A
R
A
M
pe
r
f
or
m
s
m
o
r
e
th
a
n
ot
he
r
m
e
th
od
s
ba
s
e
d
on
th
e
F
I
-
S
c
or
e
va
lu
e
p
r
oduc
e
d
a
t
m
or
e
th
a
n
90
pe
r
c
e
nt
a
nd
th
e
ha
m
m
in
g
lo
s
s
a
t
le
s
s
th
a
n
10
pe
r
c
e
nt
.
T
hi
s
r
e
s
ul
t
s
ho
w
s
th
a
t
th
is
c
om
bi
na
ti
on
c
a
n
im
pr
ove
t
he
pe
r
f
or
m
a
nc
e
of
t
he
M
L
C
m
ode
l
on da
ta
c
ont
a
in
in
g c
la
s
s
i
m
ba
la
n
c
e
.
T
a
bl
e
3
. P
os
t
hoc
te
s
t
f
or
i
nt
e
r
a
c
ti
ons
of
M
L
C
m
e
th
ods
w
it
h da
t
a
s
e
pa
r
a
ti
on me
th
ods
D
a
t
a
s
e
pa
r
a
t
i
on
M
L
C
m
e
t
hod
s
F1
-
s
c
or
e
F1
-
s
c
or
e
gr
oup
M
L
C
m
e
t
hod
s
H
a
m
m
i
ng l
os
s
H
a
m
m
i
ng l
os
s
gr
oup
R
a
ndom
ML
-
A
R
A
M
0.912
A
ML
-
A
R
A
M
0.093
A
R
a
kE
l
d
0.535
B
L
a
be
l
pow
e
r
s
e
t
0.154
B
L
a
be
l
pow
e
r
s
e
t
0.527
B
B
i
na
r
y r
e
l
e
va
nc
e
0.165
C
ML
-
kN
N
0.512
C
ML
-
kN
N
0.203
D
L
a
be
l
s
p
a
c
e
pa
r
t
i
t
i
oni
ng
0.459
D
L
a
be
l
s
p
a
c
e
pa
r
t
i
t
i
oni
ng
0.235
E
B
i
na
r
y r
e
l
e
va
nc
e
0.456
D
R
a
kE
l
d
0.253
F
S
t
r
a
t
i
f
i
e
d
ML
-
A
R
A
M
0.955
A
L
a
be
l
pow
e
r
s
e
t
0.063
A
L
a
be
l
pow
e
r
s
e
t
0.833
B
ML
-
A
R
A
M
0.074
B
ML
-
kN
N
0.739
C
ML
-
kN
N
0.104
C
B
i
na
r
y r
e
l
e
va
nc
e
0.503
D
B
i
na
r
y r
e
l
e
va
nc
e
0.149
D
R
a
kE
l
d
0.500
D
R
a
kE
l
d
0.224
E
L
a
be
l
s
p
a
c
e
pa
r
t
i
t
i
oni
ng
0.416
E
L
a
be
l
s
p
a
c
e
pa
r
t
i
t
i
oni
ng
0.236
F
3.4. F
e
at
u
r
e
i
m
p
or
t
an
c
e
U
s
in
g
th
e
M
L
-
A
R
A
M
m
e
th
od
c
om
bi
ne
d
w
it
h
th
e
s
tr
a
ti
f
ic
a
ti
on
m
e
th
od
a
s
th
e
be
s
t
m
ode
l
in
th
e
pr
e
vi
ous
s
ta
ge
,
w
e
c
a
lc
ul
a
t
e
d
th
e
im
por
ta
nc
e
va
lu
e
s
of
61
e
xpl
a
na
to
r
y
va
r
ia
bl
e
s
(
f
e
a
tu
r
e
s
)
.
W
e
di
s
c
ove
r
e
d
th
a
t
m
a
ny
e
xpl
a
na
to
r
y
va
r
ia
bl
e
s
ha
ve
ne
ga
ti
ve
f
e
a
tu
r
e
im
por
t
a
nc
e
va
lu
e
s
.
T
hi
s
m
e
a
ns
th
a
t
th
e
pr
e
s
e
nc
e
of
th
e
s
e
e
xpl
a
na
to
r
y
va
r
ia
bl
e
s
ha
s
no
s
ig
ni
f
ic
a
nt
im
pa
c
t
on
th
e
pr
e
di
c
t
io
n
a
c
c
ur
a
c
y
of
th
e
M
L
C
m
ode
l;
th
us
,
it
c
a
n
c
om
pl
ic
a
te
th
e
m
ode
l
a
nd
in
c
r
e
a
s
e
th
e
r
is
k
of
ove
r
f
it
ti
ng.
T
o
s
i
m
pl
if
y
th
e
M
L
C
m
ode
l
of
th
e
r
is
k
pr
o
f
il
e
,
w
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
I
m
pr
ov
in
g t
he
r
is
k
p
r
of
il
e
of
I
ndone
s
ia
n e
nt
e
r
pr
i
s
e
t
ax
pay
e
r
s
u
s
in
g m
ul
ti
la
be
l
…
(
T
e
guh P
r
as
e
ty
o)
4331
de
c
id
e
d
to
pe
r
f
or
m
th
e
f
e
a
tu
r
e
s
e
le
c
ti
on
s
ta
ge
,
c
ons
id
e
r
in
g
th
e
e
xc
e
ll
e
nt
pr
e
di
c
ti
on
a
c
c
ur
a
c
y,
s
pe
c
if
ic
a
ll
y
th
e
F1
-
s
c
or
e
,
w
hi
c
h
r
e
m
a
in
s
a
t
0.87,
a
nd
h
a
m
m
in
g
lo
s
s
,
w
hi
c
h
is
a
t
0.12.
T
a
bl
e
4
s
um
m
a
r
iz
e
s
20
e
xpl
a
na
to
r
y
va
r
ia
bl
e
s
(
f
e
a
tu
r
e
s
)
f
or
th
e
f
e
a
tu
r
e
s
e
le
c
ti
on
r
e
s
ul
ts
,
m
os
t
of
w
hi
c
h
a
r
e
l
oc
a
l
(
L
)
c
a
te
gor
y
r
is
k
e
le
m
e
nt
s
.
T
he
to
p t
hr
e
e
f
e
a
tu
r
e
c
ode
s
a
r
e
c
om
pa
ny
a
ge
(
R2
)
,
va
lu
e
of
a
dm
in
is
tr
a
ti
ve
s
a
nc
ti
ons
i
n t
he
f
or
m
of
f
in
e
s
(
O9
)
, a
nd
pr
oc
e
dur
e
s
f
or
im
por
ti
ng
a
nd
r
e
le
a
s
in
g
f
in
is
he
d
goods
(
L
8)
.
T
he
hi
ghe
r
th
e
f
e
a
tu
r
e
im
por
ta
nc
e
va
lu
e
,
th
e
gr
e
a
te
r
it
s
in
f
lu
e
nc
e
in
de
te
c
ti
ng
ta
x
a
voi
da
n
c
e
a
nd
ta
x
e
v
a
s
io
n.
O
ur
f
in
di
ngs
on
th
e
s
e
20
e
xpl
a
na
to
r
y
va
r
ia
bl
e
s
(
f
e
a
tu
r
e
s
)
a
r
e
th
e
r
e
s
ul
t
of
id
e
nt
if
yi
ng
r
is
k
e
le
m
e
nt
f
a
c
to
r
s
th
a
t
e
nc
our
a
ge
a
t
a
xpa
ye
r
e
nt
it
y
to
e
nga
g
e
in
ta
x
a
voi
da
nc
e
or
t
a
x e
va
s
io
n us
in
g f
e
a
tu
r
e
i
m
por
ta
nc
e
va
lu
e
s
.
T
a
bl
e
4
. S
um
m
a
r
y of
f
e
a
tu
r
e
s
e
le
c
ti
on r
e
s
ul
ts
N
o.
F
e
a
t
ur
e
i
m
por
t
a
nc
e
C
ode
N
o.
F
e
a
t
ur
e
i
m
por
t
a
nc
e
C
ode
N
o.
F
e
a
t
ur
e
i
m
por
t
a
nc
e
C
ode
1.
0
.
0722
R2
8.
0
.
0425
L
13
15.
0
.
0362
L
11
2.
0
.
0652
O9
9.
0
.
0414
R1
16.
0
.
0347
R
10
3.
0
.
0474
L8
10.
0
.
0400
L
23
17.
0
.
0344
L
29
4.
0
.
0465
L
28
11.
0
.
0395
O3
18.
0
.
0337
L
17
5.
0
.
0453
R
22
12.
0
.
0390
R8
19.
0
.
0332
O4
6.
0
.
0445
L9
13.
0
.
0382
L
12
20.
0
.
0362
O6
7.
0
.
0441
R
17
14.
0
.
0378
O1
4.
C
O
N
C
L
U
S
I
O
N
T
hi
s
s
tu
dy
s
uc
c
e
s
s
f
ul
ly
id
e
nt
if
ie
d
th
e
be
s
t
M
L
C
m
ode
l
f
o
r
de
t
e
c
ti
ng
ta
x
a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
r
is
ks
in
I
ndone
s
ia
.
T
he
be
s
t
m
e
th
od
is
th
e
M
L
-
A
R
A
M
m
e
th
o
d
ba
s
e
d
on
de
e
p
le
a
r
ni
ng,
w
hi
c
h
ha
s
e
xc
e
ll
e
nt
m
ode
l
pe
r
f
or
m
a
nc
e
a
nd
out
pe
r
f
or
m
s
to
ot
he
r
m
e
th
od
s
(
F
1
-
s
c
o
r
e
of
95.5%
a
nd
h
a
m
m
in
g
lo
s
s
of
7.4%
)
.
T
h
e
M
L
C
m
ode
l
us
e
d
in
th
is
s
tu
dy
w
il
l
im
pr
ove
th
e
a
c
c
ur
a
c
y
of
ta
x a
voi
da
nc
e
a
nd
ta
x
e
va
s
io
n
de
te
c
ti
on
in
th
e
r
is
k
pr
of
il
e
m
e
c
ha
ni
s
m
.
F
ur
th
e
r
m
or
e
,
w
e
di
s
c
ove
r
e
d
c
ondi
ti
ons
w
h
e
r
e
th
e
da
ta
c
ont
a
in
e
d
c
la
s
s
im
ba
la
nc
e
,
w
hi
c
h
t
he
s
tr
a
ti
f
ie
d
da
ta
s
e
p
a
r
a
ti
on
m
e
th
od
(
m
ul
ti
la
be
l
s
tr
a
ti
f
ie
d s
huf
f
le
s
pl
it
)
e
f
f
e
c
ti
ve
ly
a
ddr
e
s
s
e
d.
T
hi
s
is
a
n
in
put
f
or
f
ut
ur
e
s
tu
dy,
s
o
r
e
s
e
a
r
c
he
r
s
c
a
n
th
in
k a
bout
us
in
g
a
s
tr
a
ti
f
ie
d
da
ta
s
e
p
a
r
a
ti
on
m
e
th
od
to
r
e
duc
e
th
e
im
pa
c
t
of
c
la
s
s
im
ba
la
nc
e
in
th
e
M
L
C
m
e
th
od.
W
e
a
ls
o
id
e
nt
if
ie
d
r
is
k
e
le
m
e
nt
f
a
c
to
r
s
th
a
t
e
n
c
our
a
ge
ta
xp
a
ye
r
e
nt
it
ie
s
to
e
nga
ge
in
ta
x
a
voi
da
nc
e
or
ta
x
e
va
s
io
n
us
in
g
f
e
a
tu
r
e
im
por
ta
nc
e
va
lu
e
s
.
T
he
s
e
f
a
c
to
r
s
w
il
l
be
u
s
e
d
to
im
pr
ove
th
e
r
is
k
pr
of
il
e
m
e
c
ha
ni
s
m
a
nd
opt
im
iz
e
ta
x
r
e
ve
nu
e
s
i
n
I
ndon
e
s
ia
.
W
e
r
e
c
ogni
z
e
th
a
t
th
is
s
tu
dy
h
a
s
li
m
it
a
ti
ons
, s
pe
c
if
ic
a
ll
y t
ha
t
th
e
c
ha
r
a
c
te
r
is
ti
c
s
of
t
a
xpa
y
e
r
be
ha
vi
or
w
il
l
a
lwa
ys
c
ha
nge
dyna
m
ic
a
ll
y r
e
s
pons
e
to
c
ha
nge
s
in
ta
x
r
e
gul
a
ti
ons
.
T
o
a
ddr
e
s
s
th
is
,
w
e
r
e
c
om
m
e
nd
th
a
t
th
e
a
na
ly
s
is
be
r
e
pe
a
te
d
in
th
e
f
ut
ur
e
to
pr
o
duc
e
va
li
d
c
onc
lu
s
io
ns
in
im
pr
ovi
ng
th
e
r
is
k
pr
of
il
e
,
r
e
s
ul
ti
ng
in
a
m
or
e
r
e
le
va
nt
pr
of
il
e
.
W
e
a
ls
o
r
e
c
om
m
e
nd
c
onduc
ti
ng
s
im
il
a
r
s
tu
dy
to
c
om
pa
r
e
m
e
th
ods
f
or
de
a
li
ng
w
it
h
c
la
s
s
im
ba
la
nc
e
in
M
L
C
da
ta
a
nd
c
om
pa
r
in
g
ot
he
r
f
e
a
tu
r
e
im
por
ta
nc
e
m
e
th
ods
,
s
uc
h
a
s
th
e
s
ha
pl
e
y
a
ddi
ti
ve
e
xpl
a
na
ti
on
s
(
S
H
A
P
)
m
e
th
od
to
le
a
r
n
m
or
e
in
f
or
m
a
ti
on
a
bout
th
e
lo
c
a
l
in
f
lu
e
nc
e
of
e
a
c
h
e
xpl
a
na
to
r
y
va
r
ia
bl
e
(
f
e
a
tu
r
e
s
)
on
th
e
p
r
e
di
c
ti
on
a
c
c
ur
a
c
y of
t
he
M
L
C
m
ode
l.
A
C
K
N
O
WL
E
D
G
E
M
E
N
T
S
T
he
a
ut
hor
w
oul
d
li
ke
to
th
a
nk
th
os
e
w
ho
c
ont
r
ib
ut
e
d
to
th
is
s
tu
dy,
in
c
lu
di
ng
th
e
D
e
pa
r
tm
e
nt
of
S
ta
ti
s
ti
c
s
,
F
a
c
ul
ty
of
M
a
th
e
m
a
ti
c
s
a
nd
N
a
tu
r
a
l
S
c
ie
nc
e
s
,
I
P
B
U
ni
ve
r
s
it
y
f
or
f
a
c
il
it
a
ti
ng
th
e
s
tu
dy,
th
e
I
ndone
s
ia
E
ndowme
nt
F
und
f
or
E
duc
a
ti
on
A
ge
nc
y
(
L
P
D
P
)
N
o.
0000564/F
O
R
/M
/AS
N
-
2021
(
202205110209116)
f
or
f
undi
ng t
he
s
tu
dy
, a
nd t
he
I
ndone
s
ia
n M
in
i
s
tr
y of
F
in
a
nc
e
f
or
pr
ovi
di
ng t
he
s
upe
r
vi
s
io
n a
nd da
ta
s
e
t
s
.
R
E
F
E
R
E
N
C
E
S
[
1]
A
.
H
a
j
a
w
i
ya
h,
T
.
S
ur
ya
r
i
ni
,
K
i
s
w
a
nt
o,
a
nd
T
.
T
a
r
m
udj
i
,
“
A
na
l
ys
i
s
of
a
t
a
x
a
m
ne
s
t
y’
s
e
f
f
e
c
t
i
ve
ne
s
s
i
n
I
ndone
s
i
a
,”
J
our
nal
o
f
I
nt
e
r
nat
i
onal
A
c
c
ount
i
ng, A
udi
t
i
ng and T
ax
at
i
on
, vol
. 44, 2021, do
i
:
10.1016/
j
.
i
nt
a
c
c
a
udt
a
x.2021.100415.
[
2]
R
.
T
.
S
i
na
ga
,
E
.
E
va
na
,
a
nd
F
.
D
ha
r
m
a
,
“
R
e
s
e
a
r
c
h
of
t
a
x
A
voi
da
nc
e
i
n
I
ndone
s
i
a
:
A
bi
bl
i
ogr
a
phi
c
s
t
u
dy,”
A
s
i
an
J
our
nal
of
E
c
onom
i
c
s
, B
us
i
ne
s
s
and A
c
c
ount
i
ng
, vol
. 24, no.
8
, pp. 53
–
67, 2022, doi
:
10.9
734/
a
j
e
ba
/
2022/
v22i
830587.
[
3]
M
.
M
.
P
os
t
e
a
,
“
T
he
or
e
t
i
c
a
l
a
nd
m
e
t
hodol
ogi
c
a
l
a
ppr
oa
c
he
s
on
t
a
x
e
va
s
i
on,
”
D
I
E
M
:
D
ubr
ov
ni
k
I
nt
e
r
nat
i
onal
E
c
onom
i
c
M
e
e
t
i
ng
,
vol
. 6, no. 1, pp. 183
–
190, 2021, doi
:
10.17818/
di
e
m
/
2021/
1.19.
[
4]
N
.
A
l
s
a
dha
n,
“
A
m
ul
t
i
-
m
odul
e
m
a
c
hi
ne
l
e
a
r
ni
ng
a
ppr
oa
c
h
t
o
de
t
e
c
t
t
a
x
f
r
a
ud,
”
C
om
put
e
r
Sy
s
t
e
m
s
Sc
i
e
nc
e
and
E
ngi
ne
e
r
i
ng
,
vol
.
46, no. 1, pp. 241
–
253, 2023, doi
:
10.32604/
c
s
s
e
.2023.033375.
[
5]
B
.
F
.
M
ur
or
unkw
e
r
e
,
D
.
H
a
ught
on,
J
.
N
z
a
ba
ni
t
a
,
F
.
K
i
pkoge
i
,
a
nd
I
.
K
a
ba
n
o,
“
P
r
e
di
c
t
i
ng
t
a
x
f
r
a
ud
us
i
ng
s
upe
r
vi
s
e
d
m
a
c
hi
ne
l
e
a
r
ni
ng
a
ppr
oa
c
h,”
A
f
r
i
c
an
J
our
nal
of
Sc
i
e
nc
e
,
T
e
c
hnol
ogy
,
I
nnov
at
i
on
and
D
e
v
e
l
opm
e
nt
,
vol
.
15,
no.
6,
pp.
731
–
742,
2023,
do
i
:
10.1080/
20421338.2023.2187930.
[
6]
R
.
A
.
R
a
hm
a
n,
S
.
M
a
s
r
om
,
N
.
O
m
a
r
,
a
nd
M
.
Z
a
ka
r
i
a
,
“
A
n
a
ppl
i
c
a
t
i
on
of
m
a
c
hi
ne
l
e
a
r
ni
ng
on
c
or
po
r
a
t
e
t
a
x
a
voi
da
nc
e
de
t
e
c
t
i
on
m
ode
l
,”
I
A
E
S
I
nt
e
r
nat
i
onal
J
our
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
(
I
J
-
A
I
)
,
vol
.
9,
no.
4,
pp.
721
–
725,
2020,
doi
:
10.11591/
i
j
a
i
.v9.i
4.pp721
-
725.
[
7]
C
.
C
.
A
gga
r
w
a
l
,
D
at
a
c
l
as
s
i
f
i
c
at
i
on:
al
gor
i
t
hm
s
and
appl
i
c
at
i
on
s
,
B
oc
a
R
a
t
on,
F
l
or
i
da
:
C
ha
pm
a
n
a
nd
H
a
l
l
/
C
R
C
,
2014
,
doi
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
. 13, No. 4, D
e
c
e
m
be
r
2024:
4323
-
4333
4332
10.1201/
b17320.
[
8]
E
.
K
.
Y
.
Y
a
pp,
X
.
L
i
,
W
.
F
.
L
u,
a
nd
P
.
S
.
T
a
n,
“
C
om
pa
r
i
s
on
of
ba
s
e
c
l
a
s
s
i
f
i
e
r
s
f
or
m
ul
t
i
-
l
a
be
l
l
e
a
r
ni
ng,”
N
e
ur
oc
om
put
i
ng
,
vol
.
394
,
pp. 51
–
60, 2020, doi
:
10.1016/
j
.ne
uc
om
.2020.01.102.
[
9]
D
.
S
ong,
A
.
V
ol
d,
K
.
M
a
da
n,
a
nd
F
.
S
c
hi
l
de
r
,
“
M
ul
t
i
-
l
a
be
l
l
e
ga
l
doc
um
e
nt
c
l
a
s
s
i
f
i
c
a
t
i
on:
A
de
e
p
l
e
a
r
ni
ng
-
ba
s
e
d
a
ppr
oa
c
h
w
i
t
h
l
a
be
l
-
a
t
t
e
nt
i
on a
nd dom
a
i
n
-
s
pe
c
i
f
i
c
pr
e
-
t
r
a
i
ni
ng,”
I
nf
or
m
at
i
on Sy
s
t
e
m
s
, vol
. 10
6, 2022, doi
:
10.1016/
j
.i
s
.2021.101718.
[
10]
E
.
D
e
ni
z
,
H
.
E
r
ba
y,
a
nd
M
. C
oş
a
r
, “
M
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on
of
e
-
c
om
m
e
r
c
e
c
u
s
t
om
e
r
r
e
vi
e
w
s
vi
a
m
a
c
hi
ne
l
e
a
r
n
i
ng,”
A
x
i
om
s
,
vol
.
11, no. 9, pp. 1
–
17, 2022, doi
:
10.3390/
a
xi
om
s
11090436.
[
11]
W
.
L
i
u,
H
. W
a
ng,
X
.
S
he
n, a
nd
I
.
W
.
T
s
a
ng, “
T
he
e
m
e
r
gi
ng
t
r
e
nds
of
m
ul
t
i
-
l
a
be
l
l
e
a
r
ni
ng,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
at
t
e
r
n
A
nal
y
s
i
s
and M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
, vol
. 44, no. 11, pp. 7955
–
7974, 2022, doi
:
10.1109/
T
P
A
M
I
.2021.3119334.
[
12]
D
. A
f
dha
l
, K
. W
.
A
na
nt
a
, a
nd W
.
S
. H
a
r
t
ono,
“
A
dve
r
s
e
dr
ug
r
e
a
c
t
i
ons
pr
e
di
c
t
i
on us
i
ng m
ul
t
i
-
l
a
be
l
l
i
ne
a
r
di
s
c
r
i
m
i
na
nt
a
na
l
ys
i
s
a
nd
m
ul
t
i
-
l
a
be
l
l
e
a
r
ni
ng,”
i
n
2020 I
nt
e
r
nat
i
onal
C
onf
e
r
e
n
c
e
on
A
dv
anc
e
d
C
om
put
e
r
Sc
i
e
nc
e
and I
nf
or
m
at
i
on Sy
s
t
e
m
s
,
I
C
A
C
SI
S 2020
,
O
c
t
. 2020, pp. 69
–
76
, doi
:
10.1109/
I
C
A
C
S
I
S
51025.2020.9263166.
[
13]
L
.
U
w
i
ne
z
a
,
H
.
-
G
.
K
i
m
,
C
.
K
.
K
i
m
,
B
.
K
i
m
,
a
nd
J
.
-
Y
.
K
i
m
,
“
A
c
c
ur
a
c
y
a
s
s
e
s
s
m
e
nt
of
t
ypi
c
a
l
m
e
t
e
or
ol
ogi
c
a
l
y
e
a
r
da
t
a
f
or
a
phot
ovol
t
a
i
c
s
ys
t
e
m
u
s
i
ng
a
boot
s
t
r
a
p
m
e
t
h
od,”
J
our
nal
of
t
he
K
o
r
e
an
Sol
a
r
E
n
e
r
gy
Soc
i
e
t
y
,
vol
.
41,
no.
4,
pp.
115
–
129,
2021,
doi
:
10.7836/
ks
e
s
.2021.41.4.115.
[
14]
E
.
Z
i
vot
,
I
nt
r
oduc
t
i
on
t
o
c
om
put
at
i
ona
l
f
i
nanc
e
and
f
i
nanc
i
al
e
c
onom
e
t
r
i
c
s
,
2
011.
A
c
c
e
s
s
e
d:
O
c
t
.
18,
2023.
[
O
nl
i
ne
]
.
A
va
i
l
a
bl
e
:
ht
t
ps
:
/
/
bookdow
n.or
g/
c
om
pf
i
ne
z
book/
i
nt
r
oc
om
pf
i
n
r
/
[
15]
F
.
C
ha
r
t
e
,
“
A
c
om
pr
e
he
ns
i
ve
a
nd
di
da
c
t
i
c
r
e
vi
e
w
on
m
ul
t
i
l
a
be
l
l
e
a
r
ni
ng
s
of
t
w
a
r
e
t
oo
l
s
,”
I
E
E
E
A
c
c
e
s
s
,
vol
.
8,
pp.
50330
–
50354,
2020, doi
:
10.1109/
A
C
C
E
S
S
.2020.2979787.
[
16]
F
.
C
ha
r
t
e
a
nd
F
.
D
.
C
h
a
r
t
e
,
“
W
or
ki
ng
w
i
t
h
m
ul
t
i
l
a
be
l
da
t
a
s
e
t
s
i
n
r
:
t
he
m
l
dr
p
a
c
k
a
ge
,”
T
he
R
J
ou
r
nal
,
no.
7,
vol
.
2,
pp.
149
-
162,
201
5
.
[
17]
A
.
N
.
T
a
r
e
ke
gn,
M
.
G
i
a
c
obi
ni
,
a
nd
K
.
M
i
c
ha
l
a
k,
“
A
r
e
vi
e
w
o
f
m
e
t
hods
f
or
i
m
ba
l
a
nc
e
d
m
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on,”
P
at
t
e
r
n
R
e
c
ogni
t
i
on
, vol
. 118, 2021, doi
:
10.1016/
j
.pa
t
c
og.2021.107965.
[
18]
Y
.
E
.
K
ur
ni
a
w
a
t
i
a
nd
Y
.
D
.
P
r
a
bow
o,
“
M
ode
l
opt
i
m
i
s
a
t
i
on
of
c
l
a
s
s
i
m
ba
l
a
nc
e
d
l
e
a
r
ni
ng
us
i
ng
e
ns
e
m
bl
e
c
l
a
s
s
i
f
i
e
r
on
ove
r
-
s
a
m
pl
i
ng
da
t
a
,”
I
A
E
S
I
nt
e
r
nat
i
onal
J
our
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
(
I
J
-
A
I
)
,
vol
.
11,
no.
1,
pp.
276
–
283,
2022,
doi
:
10.11591/
i
j
a
i
.v11.i
1.pp276
-
283.
[
19]
B
. L
i
u, K
. B
l
e
ka
s
, a
nd G
. T
s
oum
a
ka
s
, “
M
ul
t
i
-
l
a
be
l
s
a
m
pl
i
ng ba
s
e
d on l
oc
a
l
l
a
b
e
l
i
m
ba
l
a
nc
e
,”
P
at
t
e
r
n
R
e
c
ogni
t
i
on
, vol
. 122, 2022,
doi
:
10.1016/
j
.pa
t
c
og.2021.108294.
[
20]
K
.
S
e
c
hi
di
s
,
G
.
T
s
oum
a
ka
s
,
a
nd
I
.
V
l
a
ha
va
s
,
“
O
n
t
he
s
t
r
a
t
i
f
i
c
a
t
i
on
of
m
ul
t
i
-
l
a
be
l
da
t
a
,”
M
ac
hi
ne
L
e
a
r
ni
ng
and
K
now
l
e
dg
e
D
i
s
c
ov
e
r
y
i
n D
at
abas
e
s
, pp. 145
–
158, 2011, doi
:
10.1007/
978
-
3
-
642
-
23808
-
6_
10.
[
21]
F
.
C
ha
r
t
e
,
A
.
R
i
ve
r
a
,
M
.
J
.
D
.
J
e
s
us
,
a
nd
F
.
H
e
r
r
e
r
a
,
“
C
onc
ur
r
e
nc
e
a
m
ong
i
m
ba
l
a
nc
e
d
l
a
be
l
s
a
nd
i
t
s
i
nf
l
ue
nc
e
on
m
ul
t
i
l
a
be
l
r
e
s
a
m
pl
i
ng
a
l
gor
i
t
hm
s
,”
i
n
H
y
br
i
d
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
Sy
s
t
e
m
s
,
V
e
r
l
a
g
:
S
pr
i
nge
r
,
2014,
pp.
110
–
121
,
doi
:
10.1007/
978
-
3
-
319
-
07617
-
1_10.
[
22]
P
.
S
z
ym
a
ns
ki
a
nd
T
.
K
a
j
da
now
i
c
z
,
“
S
c
i
ki
t
-
m
ul
t
i
l
e
a
r
n:
A
pyt
hon
l
i
br
a
r
y
f
or
m
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on,”
J
our
nal
of
M
ac
hi
ne
L
e
a
r
ni
ng
R
e
s
e
ar
c
h
, vol
. 20, no. 6, pp. 1
–
22, 2019.
[
23]
F
.
H
e
r
r
e
r
a
,
F
.
C
ha
r
t
e
,
A
.
J
.
R
i
ve
r
a
,
a
nd
M
.
J
.
D
.
J
e
s
us
,
M
ul
t
i
l
abe
l
c
l
as
s
i
f
i
c
at
i
on
pr
obl
e
m
anal
y
s
i
s
,
m
e
t
r
i
c
s
and
t
e
c
hni
que
s
.
S
w
i
t
z
e
r
l
a
nd
:
S
pr
i
nge
r
, 2016
,
doi
:
10.1007/
978
-
3
-
319
-
41111
-
8
.
[
24]
S
.
G
odbol
e
a
nd
S
.
S
a
r
a
w
a
gi
,
“
D
i
s
c
r
i
m
i
na
t
i
ve
m
e
t
hods
f
or
m
ul
t
i
-
l
a
be
l
e
d
c
l
a
s
s
i
f
i
c
a
t
i
on.
i
n
a
dva
nc
e
s
i
n
know
l
e
dge
di
s
c
ove
r
y
a
nd
da
t
a
m
i
ni
n
g,”
A
dv
anc
e
s
i
n K
now
l
e
dge
D
i
s
c
ov
e
r
y
and D
at
a M
i
ni
ng
,
pp. 22
–
30, 2004
, doi
:
10.1007/
978
-
3
-
540
-
24775
-
3_5
.
[
25]
F
.
L
i
,
X
. M
a
,
a
nd
Y
.
W
a
ng,
“
A
m
ul
t
i
-
l
a
be
l
m
e
t
hod
of
s
t
a
t
e
pa
r
t
i
t
i
on
a
nd
f
a
ul
t
di
a
gnos
i
s
b
a
s
e
d
on
bi
na
r
y
r
e
l
e
va
nc
e
a
l
gor
i
t
hm
,”
2020
I
E
E
E
9t
h
D
at
a
D
r
i
v
e
n
C
ont
r
ol
and
L
e
ar
ni
ng
Sy
s
t
e
m
s
C
onf
e
r
e
nc
e
(
D
D
C
L
S)
,
pp.
567
–
572,
2020,
doi
:
10.1109/
D
D
C
L
S
49620.2020.9275199.
[
26]
M
.
R
.
B
out
e
l
l
,
J
.
L
uo,
X
.
S
he
n,
a
nd
C
.
M
.
B
r
ow
n,
“
L
e
a
r
ni
ng
m
ul
t
i
-
l
a
be
l
s
c
e
ne
c
l
a
s
s
i
f
i
c
a
t
i
on,”
P
at
t
e
r
n
R
e
c
ogni
t
i
on
,
vol
.
37,
no.
9,
pp. 1757
–
1771, S
e
p. 2004, doi
:
10.1016/
j
.pa
t
c
og.2004.03.009.
[
27]
A
. S
. A
r
di
a
nt
o a
nd S
. A
di
,
“
T
he
be
s
t
pr
obl
e
m
t
r
a
ns
f
or
m
a
t
i
on m
e
t
hod i
n m
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on t
e
xt
f
or
t
he
s
i
s
a
bs
t
r
a
c
t
,”
2022 6t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
I
nf
or
m
at
i
on
T
e
c
hnol
ogy
,
I
nf
or
m
at
i
on
S
y
s
t
e
m
s
an
d
E
l
e
c
t
r
i
c
al
E
ngi
ne
e
r
i
ng
(
I
C
I
T
I
SE
E
)
,
pp.
326
–
329,
2022, doi
:
10.1109/
I
C
I
T
I
S
E
E
57756.2022.10057824.
[
28]
F
. B
e
ni
t
e
s
a
nd E
. S
a
poz
hni
kova
, “
H
A
R
A
M
:
a
hi
e
r
a
r
c
hi
c
a
l
A
R
A
M
ne
ur
a
l
ne
t
w
or
k f
or
l
a
r
ge
-
s
c
a
l
e
t
e
xt
c
l
a
s
s
i
f
i
c
a
t
i
on,”
i
n
15t
h I
E
E
E
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on D
at
a M
i
ni
ng W
or
k
s
hop, I
C
D
M
W
2015
, J
a
n. 2016, pp. 847
–
854
, doi
:
10.1109/
I
C
D
M
W
.2015.14.
[
29]
F
.
B
e
ni
t
e
s
a
nd
E
. S
a
poz
hni
kova
,
“
I
m
pr
ovi
ng
s
c
a
l
a
bi
l
i
t
y
of
A
R
T
ne
ur
a
l
ne
t
w
or
ks
,”
N
e
ur
oc
o
m
put
i
ng
,
vol
.
230,
pp.
219
–
229,
2017,
doi
:
10.1016/
j
.ne
uc
om
.2016.12.022.
[
30]
M
.
L
.
Z
ha
ng
a
nd
Z
.
H
.
Z
hou,
“
M
L
-
K
N
N
:
A
l
a
z
y
l
e
a
r
ni
ng
a
ppr
oa
c
h
t
o
m
ul
t
i
-
l
a
b
e
l
l
e
a
r
ni
ng,”
P
at
t
e
r
n
R
e
c
ogni
t
i
on
,
vol
.
40,
no.
7,
pp.
2038
–
2048, J
ul
. 2007, doi
:
10.1016/
j
.pa
t
c
og.2006.12.019.
[
31]
E
.
C
.
G
a
t
t
o,
M
.
F
e
r
r
a
ndi
n
,
a
nd
R
.
C
e
r
r
i
,
“
M
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on
w
i
t
h
l
a
be
l
c
l
us
t
e
r
s
,”
R
e
s
e
ar
c
h
Squar
e
,
2023
,
doi
:
10.21203/
r
s
.3.r
s
-
3133411/
v1
.
[
32]
J
.
M
.
M
oya
no,
E
.
L
.
G
i
ba
j
a
,
K
.
J
.
C
i
os
,
a
nd
S
.
V
e
nt
ur
a
,
“
R
e
vi
e
w
of
e
ns
e
m
bl
e
s
of
m
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
e
r
s
:
M
ode
l
s
,
e
xpe
r
i
m
e
nt
a
l
s
t
udy
a
nd pr
os
pe
c
t
s
,”
I
nf
or
m
at
i
on F
us
i
on
, vol
. 44, pp. 33
–
45, 2018, doi
:
10.1016/
j
.i
nf
f
us
.2017.12.001.
[
33]
S
.
B
a
t
e
s
,
T
.
H
a
s
t
i
e
,
a
nd
R
.
T
i
bs
hi
r
a
ni
,
“
C
r
os
s
-
va
l
i
da
t
i
on:
w
ha
t
doe
s
i
t
e
s
t
i
m
a
t
e
a
nd
how
w
e
l
l
doe
s
i
t
do
i
t
?
,”
J
our
nal
of
t
he
A
m
e
r
i
c
an
St
at
i
s
t
i
c
al
A
s
s
oc
i
at
i
on
, pp. 1
–
43, 2023, doi
:
10.1080/
01621459.2023.2197686.
[
34]
X.
-
Z
.
W
u
a
nd
Z
.
-
H
.
Z
hou,
“
A
uni
f
i
e
d
vi
e
w
of
m
ul
t
i
-
l
a
be
l
pe
r
f
or
m
a
nc
e
m
e
a
s
u
r
e
s
,”
i
n
P
r
oc
e
e
di
ngs
of
t
he
34t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on M
ac
hi
ne
L
e
ar
ni
ng
,
pp. 3780
-
3788
,
2017
.
[
35]
L
.
X
ue
,
X
.
Z
ha
ng,
W
.
J
i
a
ng,
a
nd
K
.
H
uo,
“
A
c
l
a
s
s
i
f
i
c
a
t
i
on
pe
r
f
or
m
a
nc
e
e
va
l
ua
t
i
on
m
e
a
s
ur
e
c
ons
i
de
r
i
ng
da
t
a
s
e
pa
r
a
bi
l
i
t
y,”
A
r
t
i
f
i
c
i
al
N
e
ur
al
N
e
t
w
or
k
s
and M
ac
hi
ne
L
e
ar
ni
ng
–
I
C
A
N
N
2023
,
pp. 1
-
13, 2023.
[
36]
H
.
D
a
l
i
a
ni
s
,
C
l
i
ni
c
al
t
e
x
t
m
i
ni
ng:
Se
c
ondar
y
us
e
of
e
l
e
c
t
r
oni
c
pat
i
e
nt
r
e
c
or
d
s
,
C
ha
m
:
S
pr
i
nge
r
,
2018
,
doi
:
10.1007/
978
-
3
-
319
-
78503
-
5.
[
37]
J
. R
.
-
S
a
l
a
z
a
r
,
D
at
a
s
c
i
e
nc
e
and anal
y
t
i
c
s
w
i
t
h P
y
t
hon
,
B
oc
a
R
a
t
on, F
l
or
i
da
:
C
h
a
pm
a
n a
nd H
a
l
l
/
C
R
C
, 2017.
[
38]
A
.
T
ha
r
w
a
t
,
“
C
l
a
s
s
i
f
i
c
a
t
i
on
a
s
s
e
s
s
m
e
nt
m
e
t
hods
,
”
A
ppl
i
e
d
C
o
m
put
i
ng
and
I
n
f
or
m
at
i
c
s
,
vol
.
17,
no.
1,
pp.
168
–
192,
2018,
doi
:
10.1016/
j
.a
c
i
.2018.08.003.
[
39]
M
.
R
a
pp,
E
.
L
.
M
e
n
c
í
a
,
J
.
F
ür
nkr
a
nz
,
V
.
L
.
N
guy
e
n,
a
nd
E
.
H
ül
l
e
r
m
e
i
e
r
,
“
L
e
a
r
ni
ng
gr
a
di
e
nt
boos
t
e
d
m
ul
t
i
-
l
a
be
l
c
l
a
s
s
i
f
i
c
a
t
i
on
r
ul
e
s
,”
M
ac
hi
ne
L
e
ar
ni
ng and K
now
l
e
dg
e
D
i
s
c
o
v
e
r
y
i
n D
at
aba
s
e
s
, pp. 124
–
14
0, 2021, doi
:
10.1007/
978
-
3
-
030
-
67664
-
3_8.
[
40]
H
. S
.
A
. K
ut
ubi
, “
O
n
r
a
ndom
i
z
e
d c
om
pl
e
t
e
bl
oc
k de
s
i
gn,”
I
nt
e
r
nat
i
onal
J
our
nal
of
Sc
i
e
nc
e
s
:
B
as
i
c
and A
ppl
i
e
d R
e
s
e
ar
c
h,
vol
. 53,
no.
2
,
pp. 230
-
243,
202
0
.
[
41]
C
.
M
ol
na
r
,
I
n
t
e
r
pr
e
t
a
bl
e
m
ac
h
i
n
e
l
e
ar
n
i
ng
a
gu
i
de
f
or
m
ak
i
ng
bl
ac
k
b
ox
m
ode
l
s
e
x
pl
ai
n
a
b
l
e
,
V
i
c
t
o
r
i
a
,
C
a
na
da
:
L
e
a
n
P
u
bl
i
s
h
i
ng
,
2
022
.
[
42]
D
.
A
.
O
t
c
he
r
e
,
M
.
A
boa
gye
,
M
.
A
.
A
bda
l
l
a
,
a
nd
T
.
B
.
B
oa
kye
,
“
E
nha
nc
i
ng
dr
i
l
l
i
ng
f
l
ui
d
l
os
t
-
c
i
r
c
ul
a
t
i
on
pr
e
di
c
t
i
on
us
i
ng
m
ode
l
Evaluation Warning : The document was created with Spire.PDF for Python.