I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
2
,
Apr
il
20
25
,
pp.
96
3
~
97
4
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
2
.
pp
96
3
-
97
4
963
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
A
n
e
n
h
a
n
c
e
d
c
asc
a
d
e
e
n
se
m
b
le
m
e
t
h
od
f
or
b
ig
d
at
a a
n
al
ysi
s
I
van
I
z
on
in
1
,
2
,
Ro
m
an
M
u
z
yk
a
2
,
Rom
a
n
T
k
ac
h
e
n
k
o
3
,
M
ichal
Gre
gu
s
4
,
Rom
an
Korz
h
5
,
Kyrylo
Ye
m
e
t
s
2
1
D
e
pa
r
tm
e
nt
of
C
iv
il
E
ngi
ne
e
r
in
g, S
c
hool
of
E
ngi
ne
e
r
in
g, U
ni
v
e
r
s
it
y of
B
ir
mi
ngha
m, B
ir
mi
ngha
m, U
ni
te
d K
in
gdom
2
D
e
pa
r
tm
e
nt
of
A
r
ti
f
ic
ia
l
I
nt
e
ll
ig
e
nc
e
, L
vi
v P
ol
yt
e
c
hni
c
N
a
ti
ona
l
U
ni
ve
r
s
it
y, L
vi
v, U
kr
a
in
e
3
D
e
pa
r
tm
e
nt
of
P
ubl
is
hi
ng I
nf
or
ma
ti
on T
e
c
hnol
ogi
e
s
, L
vi
v P
ol
yt
e
c
hni
c
N
a
ti
ona
l
U
ni
ve
r
s
it
y, L
vi
v, U
kr
a
in
e
4
F
a
c
ul
ty
of
M
a
na
ge
m
e
n
t,
C
ome
ni
us
U
ni
ve
r
s
it
y B
r
a
ti
s
la
va
, B
r
a
t
is
la
va
, S
lo
va
k
ia
5
D
e
pa
r
tm
e
nt
of
S
oc
ia
l
C
omm
uni
c
a
ti
on a
nd I
nf
or
ma
ti
on A
c
ti
vi
ti
e
s
, L
vi
v P
ol
yt
e
c
hni
c
N
a
ti
ona
l
U
ni
ve
r
s
it
y, L
vi
v, U
kr
a
in
e
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
J
uly
8
,
2024
R
e
vis
e
d
Oc
t
23
,
2024
Ac
c
e
pted
Nov
14
,
2024
In
t
h
e
d
i
g
i
t
al
ag
e,
t
h
e
p
r
o
l
i
ferat
i
o
n
o
f
d
a
t
a
p
res
e
n
t
s
b
o
t
h
ch
a
l
l
e
n
g
e
s
an
d
o
p
p
o
r
t
u
n
i
t
i
e
s
,
p
art
i
cu
l
arl
y
i
n
t
h
e
real
m
o
f
b
i
g
d
at
a,
w
h
i
c
h
i
s
ch
ara
ct
er
i
zed
b
y
i
t
s
v
o
l
u
me,
v
e
l
o
c
i
t
y
,
an
d
v
ar
i
et
y
.
Mach
i
n
e
l
earn
i
n
g
i
s
a
c
ru
ci
a
l
t
ech
n
o
l
o
g
y
fo
r
ex
t
rac
t
i
n
g
i
n
s
i
g
h
t
s
fro
m
t
h
es
e
v
a
s
t
d
a
t
as
e
t
s
.
A
mo
n
g
mach
i
n
e
l
ear
n
i
n
g
met
h
o
d
s
,
en
s
emb
l
e
met
h
o
d
s
,
an
d
es
p
eci
al
l
y
cas
cad
i
n
g
en
s
em
b
l
e
s
,
are
h
i
g
h
l
y
effect
i
v
e
fo
r
b
i
g
d
at
a
a
n
al
y
s
i
s
.
W
h
i
l
e
i
t
i
s
t
r
u
e
t
h
at
t
h
e
t
r
ai
n
i
n
g
p
ro
c
ed
u
res
fo
r
cas
cad
e
e
n
s
em
b
l
e
s
can
b
e
t
i
me
-
co
n
s
u
mi
n
g
a
n
d
ma
y
h
a
v
e
l
i
m
i
t
a
t
i
o
n
s
i
n
t
erms
o
f
accu
racy
,
t
h
i
s
p
ap
er
p
r
o
p
o
s
e
s
a
s
o
l
u
t
i
o
n
t
o
en
h
an
ce
t
h
ei
r
p
erfo
rman
ce.
O
u
r
met
h
o
d
i
n
v
o
l
v
e
s
u
s
i
n
g
s
t
o
ch
a
s
t
i
c
g
ra
d
i
e
n
t
d
e
s
cen
t
(
SG
D
)
cl
as
s
i
f
i
ers
,
a
n
i
mp
r
o
v
e
d
t
rai
n
i
n
g
d
at
a
s
ep
ara
t
i
o
n
al
g
o
r
i
t
h
m,
an
d
i
n
t
e
g
rat
i
n
g
p
ri
n
ci
p
al
co
mp
o
n
e
n
t
a
n
al
y
s
i
s
(PCA
)
a
t
eac
h
en
s
emb
l
e
l
ev
e
l
.
W
e
ar
e
co
n
f
i
d
e
n
t
t
h
at
t
h
e
s
e
en
h
a
n
cemen
t
s
l
ead
t
o
i
mp
r
o
v
e
d
res
u
l
t
s
a
n
d
acc
u
racy
.
T
h
e
p
ro
p
o
s
ed
a
p
p
r
o
ach
i
s
d
es
i
g
n
ed
t
o
en
h
an
ce
b
o
t
h
t
h
e
g
e
n
eral
i
zat
i
o
n
p
r
o
p
er
t
i
e
s
an
d
accu
rac
y
o
f
t
h
e
en
s
em
b
l
e
(3
%
),
w
h
i
l
e
a
l
s
o
red
u
c
i
n
g
i
t
s
t
ra
i
n
i
n
g
t
i
me.
Res
u
l
t
s
fr
o
m
mo
d
el
l
i
n
g
o
n
a
real
-
w
o
rl
d
b
i
o
med
i
cal
d
a
t
as
e
t
d
emo
n
s
t
rat
e
s
i
g
n
i
fi
can
t
red
u
c
t
i
o
n
s
i
n
t
rai
n
i
n
g
d
u
rat
i
o
n
,
i
mp
ro
v
emen
t
s
i
n
g
e
n
eral
i
zat
i
o
n
p
ro
p
ert
i
e
s
,
an
d
en
h
a
n
ce
d
accu
racy
w
h
en
c
o
mp
are
d
t
o
o
t
h
er
p
o
s
s
i
b
l
e
i
m
p
l
eme
n
t
a
t
i
o
n
s
o
f
t
h
e
en
s
em
b
l
e
.
K
e
y
w
o
r
d
s
:
B
ig
da
ta
a
na
lys
is
B
inar
y
c
las
s
if
ica
ti
on
tas
k
C
a
s
c
a
d
e
e
ns
e
mbl
e
I
mbala
nc
e
d
da
tas
e
t
Kolmogor
ov
-
Ga
bor
polynom
ial
M
a
c
hine
lea
r
ning
W
iene
r
polynom
ial
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
R
oman
M
uz
yka
De
pa
r
tm
e
nt
of
A
r
ti
f
icia
l
I
ntelli
ge
nc
e
,
L
viv
P
olyt
e
c
hnic
Na
ti
ona
l
Unive
r
s
it
y
S
.
B
a
nde
r
a
s
tr
.
,
12,
L
viv,
79013
,
Ukr
a
ine
E
mail:
r
oman.
muzyka
.
mkns
s
h.
2022@lpnu.
ua
1.
I
NT
RODU
C
T
I
ON
I
n
the
dig
it
a
l
e
r
a
,
th
e
e
x
pone
n
ti
a
l
g
r
ow
th
of
da
ta
ha
s
us
he
r
e
d
i
n
ne
w
c
ha
ll
e
n
ge
s
a
n
d
op
po
r
tun
it
ies
.
B
ig
da
ta
is
o
f
ten
c
ha
r
a
c
te
r
ize
d
by
a
mo
ng
othe
r
thi
n
gs
,
it
s
volu
me,
ve
loci
ty
,
a
n
d
va
r
ie
ty
,
wh
ich
c
a
n
pos
e
a
f
o
r
mi
da
ble
c
ha
ll
e
n
ge
to
c
on
ve
nt
iona
l
da
ta
p
r
oc
e
s
s
ing
tec
hn
iq
ue
s
.
How
e
v
e
r
,
mac
h
ine
lea
r
ni
ng
ha
s
e
me
r
ge
d
a
s
a
pi
vota
l
tec
hno
logy
i
n
a
d
dr
e
s
s
i
ng
thes
e
c
ha
l
lenge
s
,
du
e
to
it
s
a
bil
i
ty
to
a
na
lyze
a
nd
e
xt
r
a
c
t
i
ns
igh
ts
f
r
o
m
mas
s
ive
da
tas
e
ts
.
T
he
s
yne
r
gy
be
twe
e
n
m
a
c
hine
lea
r
n
ing
a
nd
bi
g
d
a
ta
p
r
oc
e
s
s
in
g
ha
s
unde
r
g
one
s
igni
f
ica
nt
de
ve
lop
ment
ove
r
ti
me
.
H
owe
ve
r
,
c
on
ve
nt
iona
l
da
t
a
p
r
oc
e
s
s
in
g
tec
hniques
ha
ve
e
nc
ou
nte
r
e
d
c
ha
ll
e
ng
e
s
i
n
ha
n
dli
ng
the
va
s
t
a
moun
t
a
nd
in
tr
ica
c
y
o
f
big
da
ta
[
1
]
,
[
2
]
.
T
h
e
d
e
ploy
ment
o
f
mac
h
ine
lea
r
ni
ng
mo
de
ls
a
t
s
c
a
le
ha
s
be
e
n
s
igni
f
ica
n
tl
y
im
pr
ove
d
by
mac
h
ine
lea
r
nin
g
a
lgo
r
i
thm
s
,
pa
r
t
icula
r
l
y
t
hos
e
that
ut
il
iz
e
pa
r
a
l
lel
c
o
mp
u
ti
ng
a
nd
dis
tr
ibu
ted
s
ys
tems
.
T
his
ha
s
a
ll
owe
d
o
r
ga
niza
t
ion
s
to
e
xt
r
a
c
t
va
l
ue
f
r
om
thei
r
da
ta
a
s
s
e
ts
m
or
e
e
f
f
ic
i
e
ntl
y
[
3]
.
T
he
a
dve
n
t
o
f
tec
hn
olo
gies
s
uc
h
a
s
a
pa
c
he
ha
doo
p
a
nd
s
pa
r
k
ha
s
ma
de
s
c
a
la
ble
a
nd
e
f
f
ic
ient
big
da
ta
p
r
oc
e
s
s
ing
f
r
a
mew
o
r
ks
a
c
c
e
s
s
ibl
e
,
the
r
e
b
y
f
a
c
i
li
tati
ng
the
de
pl
oymen
t
of
mac
hine
lea
r
n
ing
m
ode
ls
a
t
s
c
a
le
[
4]
.
M
a
c
hine
lea
r
ning
ha
s
s
e
ve
r
a
l
a
dva
ntag
e
s
f
or
pr
oc
e
s
s
ing
bi
g
da
ta
tas
ks
[
5]
.
F
ir
s
tl
y,
it
e
na
bles
pr
e
dictive
a
na
lyt
ics
by
identif
ying
pa
tt
e
r
ns
a
nd
tr
e
nds
withi
n
va
s
t
da
tas
e
ts
,
whic
h
c
a
n
f
a
c
il
it
a
te
in
f
or
med
de
c
is
ion
-
making
[
6]
.
S
e
c
ondly,
mac
hine
lea
r
ning
a
lgor
it
hms
c
a
n
a
utom
a
te
da
ta
pr
oc
e
s
s
ing
wor
kf
lows
,
r
e
duc
ing
manua
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
96
3
-
97
4
964
int
e
r
ve
nti
on
a
nd
s
tr
e
a
ml
ini
ng
ope
r
a
ti
ons
[
7]
.
T
hir
dl
y,
it
is
im
por
tant
to
note
that
the
it
e
r
a
ti
ve
na
tur
e
of
mac
hine
lea
r
ning
a
ll
ows
models
to
c
onti
nuous
ly
im
pr
ove
a
nd
a
da
pt
to
e
volvi
ng
da
tas
e
ts
,
r
e
s
ult
ing
in
e
nha
nc
e
d
a
c
c
ur
a
c
y
a
nd
pe
r
f
or
manc
e
ove
r
ti
me
[
8]
.
De
s
pit
e
it
s
pr
omi
s
ing
c
a
pa
bil
it
ies
,
mac
hine
lea
r
ning
f
or
big
da
ta
pr
oc
e
s
s
ing
is
not
without
c
ha
ll
e
nge
s
[
9
]
.
T
he
main
hu
r
dle
is
th
a
t
lar
ge
da
ta
s
e
ts
a
r
e
c
ompl
e
x
to
mana
ge
a
nd
e
f
f
e
c
ti
ve
ly
pr
oc
e
s
s
[
10]
.
E
ns
ur
ing
s
c
a
labili
ty,
f
a
ult
tol
e
r
a
nc
e
,
a
nd
r
e
s
our
c
e
opti
mi
z
a
ti
on
in
dis
tr
ibu
ted
e
nvir
onments
r
e
mains
a
n
ongoing
c
ha
ll
e
nge
.
I
n
a
ddit
ion,
da
ta
qu
a
li
ty
[
11]
a
nd
pr
e
pr
oc
e
s
s
ing
[
12]
a
r
e
c
r
it
ica
l
f
a
c
to
r
s
that
c
a
n
s
igni
f
ica
ntl
y
im
pa
c
t
th
e
pe
r
f
o
r
manc
e
of
mac
hine
le
a
r
ning
models
[
13]
,
[
14
]
.
Add
r
e
s
s
ing
is
s
ue
s
s
uc
h
a
s
mi
s
s
ing
va
lues
,
outl
ier
s
,
a
nd
da
ta
im
ba
lanc
e
s
r
e
quir
e
s
c
a
r
e
f
ul
c
ons
ider
a
ti
on
to
pr
e
ve
nt
bias
a
nd
inac
c
ur
a
c
y
in
the
a
na
lys
is
[
15]
,
[
16]
.
F
ur
ther
mo
r
e
,
the
a
c
c
ur
a
c
y
of
in
divi
dua
l
mac
hine
lea
r
ning
models
,
whic
h
is
c
r
it
ica
l
f
or
s
uc
h
tas
ks
,
is
not
a
lwa
ys
s
a
ti
s
f
a
c
tor
y.
I
n
pa
r
ti
c
ular
,
the
method
de
s
c
r
ibed
[
17]
is
highl
y
e
f
f
e
c
ti
ve
f
or
a
na
ly
z
in
g
lar
ge
da
tas
e
ts
.
How
e
ve
r
,
the
a
c
c
ur
a
c
y
of
their
s
tocha
s
ti
c
gr
a
dient
de
s
c
e
nt
(
S
GD
)
a
lgo
r
it
hm
is
not
up
to
the
mar
k.
T
o
a
ddr
e
s
s
thi
s
is
s
ue
,
I
z
onin
e
t
al
.
[
18]
inves
ti
g
a
ted
the
nonl
inea
r
e
xpa
ns
ion
of
input
s
f
or
S
GD
,
im
pleme
nti
ng
it
us
ing
dif
f
e
r
e
nt
powe
r
s
of
the
W
ie
ne
r
polynom
ial.
T
his
pa
pe
r
a
ppr
oxim
a
tes
a
tabula
r
da
tas
e
t
us
ing
dif
f
e
r
e
nt
powe
r
s
o
f
the
s
a
me
polynom
ial
.
T
h
e
de
gr
e
e
of
the
po
lynom
ial
wa
s
de
ter
mi
ne
d
us
ing
S
GD
due
to
it
s
high
s
pe
e
d.
T
he
r
e
s
ult
s
of
the
modelli
ng
s
ugge
s
t
that
incr
e
a
s
ing
the
de
gr
e
e
of
a
ppr
oxim
a
ti
on
would
im
pr
ove
the
a
c
c
ur
a
c
y
of
the
model
.
How
e
ve
r
,
t
he
dir
e
c
t
a
ppr
oxim
a
ti
on
a
ppr
oa
c
h
by
high
powe
r
s
of
thi
s
polynom
ial
may
s
igni
f
ica
ntl
y
incr
e
a
s
e
the
pr
oble
m's
dim
e
ns
ionali
ty.
I
t
is
im
po
r
tant
to
note
that
po
lynom
ial
a
ppr
oxim
a
ti
on
may
not
a
lwa
ys
be
a
pp
r
opr
iate
whe
n
the
number
of
a
tt
r
ibut
e
s
e
xc
e
e
ds
the
number
of
v
e
c
tor
s
in
a
da
tas
e
t.
He
nc
e
,
a
dir
e
c
t
a
ppr
oxim
a
ti
on
with
thi
s
polynom
ial,
e
ve
n
with
the
us
e
of
high
-
s
pe
e
d
S
GD
,
may
n
ot
be
the
mos
t
s
uit
a
ble
a
ppr
oa
c
h.
W
he
n
dis
c
us
s
ing
the
c
ompos
it
ion
of
e
ns
e
mbl
e
s
f
r
om
thes
e
methods
[
19]
,
i
t
is
im
por
tant
to
note
that
they
c
a
n
pa
r
ti
a
ll
y
a
ll
e
viate
the
a
f
or
e
mentioned
is
s
ue
s
.
I
n
pa
r
ti
c
ular
,
s
c
a
li
ng,
a
s
the
mos
t
a
c
c
ur
a
te
c
las
s
of
e
ns
e
mbl
e
methods
,
c
a
n
be
opti
mi
z
e
d
to
wor
k
e
f
f
ici
e
ntl
y
with
lar
ge
-
s
c
a
le
da
tas
e
t
s
[
20]
.
T
he
s
e
methods
pa
r
ti
ti
on
the
da
ta
or
e
mpl
oy
incr
e
menta
l
t
r
a
ini
ng
tec
hniques
,
whic
h
c
a
n
he
lp
mana
ge
a
nd
pr
oc
e
s
s
da
ta
in
dis
tr
ibut
e
d
e
nvir
onments
mor
e
e
f
f
icie
ntl
y
.
C
a
s
c
a
de
e
ns
e
mbl
e
s
c
a
n
be
c
ons
ider
e
d
mor
e
f
a
ult
-
tol
e
r
a
nt
than
in
divi
dua
l
models
a
s
they
us
e
mul
ti
ple
models
[
21]
.
T
h
is
mea
ns
that
if
one
model
f
a
il
s
or
pr
oduc
e
s
inac
c
ur
a
te
r
e
s
ult
s
,
the
e
ns
e
mbl
e
c
a
n
s
ti
ll
make
r
e
li
a
ble
p
r
e
dictions
by
a
ggr
e
ga
t
ing
output
s
f
r
om
mul
ti
ple
models
.
F
ur
t
he
r
mor
e
,
c
a
s
c
a
de
e
n
s
e
mbl
e
s
c
a
n
opti
mi
z
e
r
e
s
our
c
e
s
by
dis
t
r
ibut
ing
c
omput
a
ti
on
a
c
r
os
s
mul
ti
ple
models
or
pr
oc
e
s
s
ing
unit
s
[
22]
,
lea
ding
to
opt
im
ize
d
uti
li
z
a
ti
on
o
f
c
omp
utational
r
e
s
our
c
e
s
in
dis
tr
ibut
e
d
e
nvir
onments
.
M
or
e
ove
r
,
it
is
wor
th
noti
ng
that
e
ns
e
mbl
e
methods
,
includi
ng
c
a
s
c
a
de
e
ns
e
mbl
e
s
,
ha
ve
the
potential
to
be
r
e
s
is
tant
to
nois
y
or
im
pe
r
f
e
c
t
da
ta
.
B
y
c
ombi
ning
mul
ti
ple
m
ode
ls
that
a
r
e
tr
a
ined
on
di
f
f
e
r
e
nt
s
ubs
e
ts
or
r
e
pr
e
s
e
ntations
of
the
da
ta,
e
ns
e
mbl
e
methods
c
a
n
mi
ti
ga
te
the
im
pa
c
t
of
mi
s
s
ing
va
lues
,
outl
ier
s
,
a
nd
da
ta
im
ba
lanc
e
s
[
16]
.
C
a
s
c
a
d
e
e
ns
e
mbl
e
methods
may
not
c
ompl
e
tely
s
ol
ve
a
ll
the
c
ha
ll
e
nge
s
mentioned
pr
e
vios
ly,
but
they
c
a
n
c
e
r
tainly
he
lp
a
dd
r
e
s
s
them
by
u
ti
li
z
ing
the
di
ve
r
s
it
y
a
nd
c
oll
e
c
ti
ve
int
e
ll
igenc
e
of
mul
ti
ple
mod
e
ls
[
23]
.
I
t
is
c
r
uc
ial
to
c
a
r
e
f
ull
y
de
s
ign
a
nd
tune
c
a
s
c
a
de
e
ns
e
mbl
e
s
to
f
it
the
s
p
e
c
if
ic
c
ha
r
a
c
ter
is
ti
c
s
a
nd
r
e
q
uir
e
ments
of
the
pr
oblem
do
main
[
24]
.
W
hil
e
c
a
s
c
a
de
e
ns
e
mbl
e
s
a
r
e
c
ons
ider
e
d
the
mos
t
a
c
c
ur
a
te
c
las
s
of
e
ns
e
mbl
e
methods
,
their
hier
a
r
c
hica
l
de
c
is
ion
-
making
pr
oc
e
s
s
r
e
quir
e
s
a
lengthy
tr
a
ini
ng
pr
oc
e
dur
e
[
25]
.
T
his
tas
k
c
a
n
be
c
ome
e
ve
n
mor
e
c
ompl
e
x
whe
n
a
na
lyzing
high
-
d
im
e
ns
ional
da
tas
e
ts
[
26]
us
ing
c
ompl
e
x,
nonli
ne
a
r
mac
hine
lea
r
ning
methods
a
t
e
a
c
h
leve
l
o
f
a
de
e
p
lea
r
ning
c
a
s
c
a
de
[
27]
.
M
or
e
ove
r
,
the
methodology
e
ntails
s
e
gmenting
the
da
tas
e
t
int
o
s
e
c
ti
ons
,
whic
h
a
r
e
then
pr
oc
e
s
s
e
d
a
t
de
s
ignate
d
leve
ls
withi
n
the
c
a
s
c
a
de
s
tr
uc
tur
e
.
T
his
tec
hnique
r
e
s
tr
icts
the
e
xpos
ur
e
of
we
a
ke
r
p
r
e
dictor
s
to
the
e
nti
r
e
da
tas
e
t,
ther
e
by
r
e
duc
ing
the
a
c
c
ur
a
c
y
of
the
c
a
s
c
a
de
f
or
e
c
a
s
t
or
c
las
s
if
ica
ti
on
model
a
s
a
whole
[
28]
.
T
he
s
e
f
a
c
tor
s
c
umul
a
ti
ve
ly
a
f
f
e
c
t
the
pe
r
f
or
manc
e
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
.
Ac
c
or
ding
to
the
li
ter
a
tu
r
e
,
a
c
ompr
e
he
ns
ive
e
va
luation
of
the
pe
r
f
or
manc
e
o
f
c
a
s
c
a
de
e
ns
e
mbl
e
s
hould
take
in
to
a
c
c
ount
va
r
ious
indi
c
a
tor
s
s
uc
h
a
s
a
c
c
ur
a
c
y,
s
pe
e
d,
a
nd
ge
ne
r
a
li
z
a
ti
on
[
29
]
.
Ac
c
ur
a
c
y
(
ba
s
e
d
on
dif
f
e
r
e
nt
pe
r
f
or
manc
e
indi
c
a
tor
s
)
mea
s
ur
e
s
how
s
uc
c
e
s
s
f
ull
y
the
mac
hine
lea
r
nin
g
model
pr
e
dicts
outcome
s
c
ompar
e
d
to
the
a
c
tual
r
e
s
ult
s
.
it
's
typi
c
a
ll
y
e
xp
r
e
s
s
e
d
a
s
a
pe
r
c
e
ntage
a
nd
is
c
r
uc
ial
f
or
e
ns
ur
ing
th
e
r
e
li
a
bil
it
y
of
ins
ight
s
de
r
ived
f
r
om
big
da
ta.
T
r
a
ini
ng
ti
me
mea
s
ur
e
s
the
dur
a
ti
on
r
e
quir
e
d
to
tr
a
in
a
mac
hine
lea
r
ning
model
on
a
given
da
tas
e
t
[
30
]
,
while
ge
ne
r
a
li
z
a
ti
on
mea
s
ur
e
s
it
s
a
bil
it
y
to
pe
r
f
or
m
we
ll
on
uns
e
e
n
da
ta.
I
t
is
im
por
tant
to
c
ons
ider
a
ll
thes
e
in
dica
tor
s
in
c
ombi
na
ti
on.
Dudz
ik
e
t
al.
[
27]
de
v
e
loped
a
c
a
s
c
a
de
e
ns
e
mbl
e
ba
s
e
d
on
s
uppor
t
ve
c
tor
mac
hines
(
S
VM
s
)
.
T
he
S
VM
e
ns
e
mbl
e
wa
s
c
ompos
e
d
us
ing
a
n
e
volut
ionar
y
a
lgor
it
hm
pr
opos
e
d
by
the
a
uthor
s
to
opti
mi
z
e
the
hype
r
pa
r
a
mete
r
s
of
the
mac
hine
lea
r
n
ing
method
unde
r
lyi
ng
the
c
a
s
c
a
de
e
ns
e
mbl
e
.
T
he
pr
o
pos
e
d
a
ppr
oa
c
h
ha
s
de
mons
tr
a
ted
high
a
c
c
ur
a
c
y.
T
he
tr
a
ini
ng
pr
oc
e
s
s
f
or
S
VM
s
[
31]
is
known
to
ha
ve
high
ti
me
a
nd
memor
y
c
ompl
e
xit
y,
whic
h
is
f
ur
ther
incr
e
a
s
e
d
by
the
opti
m
iza
ti
on
pr
oc
e
dur
e
f
o
r
e
a
c
h
S
VM
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
.
As
a
r
e
s
ult
,
the
a
c
c
ur
a
c
y
a
nd
du
r
a
ti
on
of
S
VM
s
a
r
e
li
mi
ted
,
making
thei
r
a
ppli
c
a
ti
on
c
ha
ll
e
nging.
How
e
ve
r
,
with
c
a
r
e
f
ul
c
ons
ider
a
ti
on
a
nd
e
xpe
r
ti
s
e
,
S
VM
s
c
a
n
s
ti
ll
be
a
va
luable
tool
in
c
e
r
tain
c
ontexts
.
Ac
c
or
ding
to
I
z
onin
e
t
al.
[
32
]
,
a
dis
ti
nc
t
method
wa
s
e
mpl
oye
d
by
the
a
utho
r
s
to
c
ons
tr
uc
t
a
c
a
s
c
a
de
e
ns
e
mbl
e
us
ing
s
uppor
t
ve
c
tor
r
e
gr
e
s
s
ion
(
S
VR
)
.
T
he
da
tas
e
t
wa
s
pa
r
ti
ti
one
d
int
o
e
qua
l
s
e
gments
,
with
the
number
of
s
e
gments
de
ter
mi
ning
the
c
a
s
c
a
de
's
de
pt
h.
T
he
ba
s
ic
mac
hine
lea
r
ning
method
us
e
d
wa
s
li
n
e
a
r
S
VR
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
n
e
nhanc
e
d
c
as
c
ade
e
ns
e
mble
me
thod
for
big
da
ta
analys
is
(
I
v
an
I
z
onin
)
965
T
his
a
ppr
oa
c
h
im
pr
ove
d
the
method's
s
pe
e
d,
a
lt
ho
ugh
it
may
ha
ve
de
c
r
e
a
s
e
d
the
potential
a
c
c
ur
a
c
y
of
c
a
s
c
a
de
e
ns
e
mbl
e
s
.
T
o
a
ddr
e
s
s
the
a
f
or
e
mentioned
dr
a
wba
c
k,
a
modi
f
i
c
a
ti
on
of
thi
s
s
c
he
me
wa
s
pr
opos
e
d
in
[
33]
.
At
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
,
the
a
uthor
s
ut
il
ize
d
high
-
s
pe
e
d
S
GD
a
s
a
f
unda
menta
l
mac
hine
lea
r
ning
a
lgor
it
hm.
Additi
ona
ll
y,
a
qua
dr
a
ti
c
W
iene
r
polynom
ial
wa
s
e
mpl
oye
d
f
or
the
nonli
ne
a
r
t
r
a
ns
f
or
mation
o
f
inpu
t
da
ta
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
,
whic
h
r
e
s
ult
e
d
in
a
s
ign
if
ica
nt
im
pr
ove
ment
in
the
f
or
e
c
a
s
t
a
c
c
ur
a
c
y.
M
or
e
ove
r
,
it
s
hould
be
noted
that
the
c
a
s
c
a
de
s
tr
uc
tur
e
of
th
e
method
r
e
s
ult
s
in
a
n
im
pli
c
it
a
pp
r
oxim
a
ti
on
t
hr
ough
a
high
-
de
gr
e
e
polynom
ial.
I
t
is
wo
r
th
mentioni
ng
t
ha
t
e
a
c
h
ne
w
leve
l
of
the
c
a
s
c
a
de
doubles
the
or
d
e
r
of
the
W
iene
r
polynom
ial.
How
e
ve
r
,
thi
s
incr
e
a
s
e
in
or
d
e
r
lea
ds
to
a
s
igni
f
ica
nt
e
xpa
ns
ion
o
f
the
input
da
t
a
s
pa
c
e
,
whic
h
in
tur
n
pr
olongs
the
t
r
a
ini
ng
pr
oc
e
dur
e
.
T
o
pa
r
ti
a
ll
y
c
ompens
a
te
f
o
r
thi
s
dr
a
wba
c
k,
S
GD
is
e
mpl
oye
d.
How
e
ve
r
,
it
is
wor
th
noti
ng
that
the
pr
opos
e
d
a
pp
r
oa
c
h
may
ha
ve
a
n
im
pa
c
t
on
the
ge
ne
r
a
li
z
a
ti
on
p
r
ope
r
ti
e
s
of
the
method.
T
he
r
e
f
or
e
,
it
may
be
ne
c
e
s
s
a
r
y
to
c
onduc
t
f
ur
ther
r
e
s
e
a
r
c
h
to
r
e
duc
e
the
tr
a
ini
ng
ti
me
of
the
method
while
s
im
ult
a
ne
ous
ly
im
pr
oving
it
s
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
a
nd
a
c
c
ur
a
c
y
[
34]
.
T
he
objec
ti
ve
of
thi
s
pa
pe
r
is
to
im
pr
ove
the
pe
r
f
o
r
manc
e
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
of
S
GD
c
las
s
if
ier
s
by
im
pleme
nti
ng
a
c
ombi
na
ti
on
of
a
ne
w
da
ta
pa
r
ti
ti
oning
a
lgor
it
hm
a
nd
P
C
A
a
t
e
a
c
h
leve
l
of
the
e
ns
e
mbl
e
method.
T
he
e
f
f
e
c
ti
ve
ne
s
s
of
thi
s
a
ppr
oa
c
h
is
e
v
a
luate
d
by
mea
s
ur
ing
e
nha
nc
e
ments
in
the
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
a
nd
a
c
c
ur
a
c
y
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
of
S
GD
c
las
s
if
ier
s
,
a
s
we
ll
a
s
a
s
ubs
tantial
r
e
duc
ti
on
in
the
dur
a
ti
on
of
it
s
tr
a
ini
ng
.
T
h
e
m
a
in
c
o
nt
r
ib
ut
io
ns
o
f
th
is
pa
p
e
r
a
r
e
th
e
f
ol
l
ow
in
g:
−
W
e
i
mp
r
ov
e
d
th
e
S
GD
-
b
a
s
e
d
c
a
s
c
a
de
e
ns
e
mb
le
by
j
oi
nt
l
y
u
ti
l
izi
ng
a
n
e
w
da
ta
pa
r
t
i
ti
on
in
g
a
lg
o
r
i
thm
a
nd
a
dd
i
ti
ona
l
a
p
pl
ica
t
io
n
o
f
P
C
A
a
t
e
a
c
h
le
ve
l
of
t
he
h
ie
r
a
r
c
hi
c
a
l
e
ns
e
mb
le
.
T
he
us
e
o
f
t
he
f
i
r
s
t
a
p
p
r
o
a
c
h
d
e
m
ons
t
r
a
te
d
a
s
ig
ni
f
i
c
a
nt
i
mp
r
ove
me
nt
i
n
t
he
a
c
c
u
r
a
c
y
of
t
he
e
ns
e
mb
le
me
th
od
.
T
he
u
t
il
iz
a
t
io
n
o
f
th
e
s
e
c
on
d
a
pp
r
oa
c
h
s
ig
ni
f
ica
nt
ly
r
e
d
uc
e
d
i
ts
t
r
a
in
in
g
t
i
me
.
T
h
e
c
o
mb
ine
d
us
e
o
f
bo
t
h
a
p
p
r
o
a
c
h
e
s
p
r
ov
id
e
d
a
s
u
bs
t
a
n
t
ial
e
nh
a
nc
e
m
e
n
t
i
n
t
he
pe
r
f
o
r
ma
nc
e
o
f
t
he
c
a
s
c
a
de
e
ns
e
mb
le
ba
s
e
d
o
n
two
c
r
i
ti
c
a
l
i
nd
ic
a
t
o
r
s
;
−
W
e
i
mp
r
o
ve
d
t
he
t
r
a
in
i
ng
a
nd
a
pp
li
c
a
ti
on
p
r
oc
e
d
u
r
e
s
o
f
the
c
a
s
c
a
de
e
ns
e
mb
le
th
r
o
ug
h
t
he
c
omb
i
ne
d
i
m
pl
e
m
e
n
ta
ti
on
of
bo
th
a
p
p
r
oa
c
he
s
a
s
m
e
n
ti
on
e
d
in
t
he
f
i
r
s
t
s
c
ien
t
if
ic
c
o
nt
r
ib
ut
io
n
o
f
th
is
p
a
pe
r
,
im
p
r
o
v
ing
i
ts
pe
r
f
o
r
ma
nc
e
in
te
r
ms
o
f
a
c
c
u
r
a
c
y
a
n
d
t
r
a
i
ni
ng
ti
m
e
w
he
n
s
o
lv
in
g
c
las
s
i
f
i
c
a
t
i
on
tas
ks
,
pa
r
ti
c
u
la
r
ly
in
t
he
a
na
lys
is
o
f
la
r
ge
d
a
t
a
s
e
ts
;
−
W
e
ha
ve
de
mo
ns
t
r
a
te
d
a
s
ig
ni
f
ica
n
t
e
n
ha
nc
e
me
nt
in
t
he
pe
r
f
o
r
ma
nc
e
o
f
t
he
c
a
s
c
a
de
e
ns
e
mb
le
(
tr
a
i
n
in
g
ti
m
e
,
g
e
n
e
r
a
l
iz
a
t
io
n
p
r
ope
r
t
ies
)
c
om
pa
r
e
d
t
o
o
th
e
r
p
os
s
i
ble
i
mp
le
men
ta
ti
on
s
.
T
he
pa
pe
r
is
s
tr
uc
tu
r
e
d
a
s
f
oll
ows
:
in
s
e
c
ti
on
2,
the
e
nha
nc
e
ments
made
to
the
c
a
s
c
a
de
e
ns
e
mbl
e
method
a
r
e
e
xplaine
d,
including
the
im
p
leme
ntation
of
a
nove
l
t
r
a
ini
ng
da
ta
pa
r
t
it
ioni
ng
a
lgor
it
hm
a
nd
the
int
e
gr
a
ti
on
of
pr
incipa
l
c
omponent
a
na
lys
is
(
P
C
A)
a
t
e
a
c
h
leve
l.
T
he
r
e
s
ult
s
obtaine
d
f
r
om
the
a
ppli
c
a
ti
on
of
the
im
pr
ove
d
c
a
s
c
a
de
e
n
s
e
mbl
e
method
a
r
e
pr
e
s
e
nted
in
s
e
c
ti
on
3,
a
nd
the
im
pli
c
a
ti
ons
of
the
f
ind
ings
a
r
e
dis
c
us
s
e
d.
F
inally,
s
e
c
ti
on
4
s
umm
a
r
ize
s
the
ke
y
f
i
ndings
a
nd
c
ontr
ibut
ions
of
the
s
tudy
.
2.
AN
I
M
P
ROVE
D
CA
S
CA
DE
E
NSE
M
B
L
E
M
E
T
HO
D
T
h
e
c
a
s
c
a
d
e
e
n
s
e
m
b
l
e
i
m
p
r
o
v
e
d
i
n
t
h
i
s
p
a
p
e
r
i
s
b
a
s
e
d
o
n
[
3
3
]
.
A
s
p
r
e
v
i
o
u
s
l
y
m
e
n
t
i
o
n
e
d
,
I
z
o
n
i
n
e
t
a
l
.
[
3
3
]
p
r
op
os
e
d
a
h
ie
r
a
r
c
hi
c
a
l
c
las
s
if
ie
r
t
ha
t
us
e
s
a
hi
gh
-
s
pe
e
d
S
G
D
q
ua
d
r
a
t
ic
W
i
e
n
e
r
p
ol
yn
om
ia
l
f
o
r
no
n
li
ne
a
r
t
r
a
ns
f
o
r
ma
t
io
n
o
f
th
e
in
pu
t
d
a
t
a
a
t
e
a
c
h
le
ve
l
o
f
t
he
c
a
s
c
a
de
.
T
he
t
r
a
in
in
g
d
a
t
a
s
e
t
is
d
iv
id
e
d
in
t
o
e
q
ua
l
pa
r
ts
,
a
nd
t
he
n
um
be
r
o
f
pa
r
t
s
d
e
t
e
r
m
in
e
s
the
n
um
be
r
o
f
c
a
s
c
a
de
l
e
ve
ls
.
Ho
we
ve
r
,
it
s
h
ou
ld
b
e
n
o
ted
t
ha
t
th
e
e
xis
t
ing
m
e
t
ho
d
h
a
s
tw
o
d
r
a
wba
c
ks
.
O
ne
is
the
f
o
r
ma
ti
on
o
f
r
a
nd
om
s
ubs
a
m
pl
e
s
o
f
th
e
s
a
m
e
s
ize
(
w
i
th
ou
t
r
e
pe
ti
t
io
ns
)
f
o
r
e
a
c
h
l
e
v
e
l
o
f
t
he
c
a
s
c
a
de
.
T
h
is
m
a
y
r
e
s
ul
t
in
we
a
k
r
e
gr
e
s
s
or
s
r
e
c
e
iv
i
ng
on
ly
a
s
m
a
l
l
p
o
r
t
io
n
of
t
he
us
e
f
u
l
i
n
f
o
r
ma
ti
on
f
o
r
a
na
lys
is
,
wh
ic
h
c
o
ul
d
po
te
nt
ia
ll
y
r
e
d
uc
e
the
a
c
c
ur
a
c
y
o
f
t
he
c
a
s
c
a
d
e
a
s
a
wh
ole
.
S
e
c
on
d
,
t
he
us
e
of
a
no
n
li
ne
a
r
e
xp
a
ns
io
n
s
c
h
e
m
e
f
o
r
t
he
p
r
ob
le
m
i
np
uts
ba
s
e
d
o
n
th
e
q
ua
d
r
a
t
ic
W
ie
ne
r
p
ol
yn
om
ia
l
r
e
p
r
e
s
e
n
ts
a
no
th
e
r
p
o
ten
t
ia
l
d
is
a
dv
a
n
ta
ge
.
W
h
il
e
th
is
a
p
pr
oa
c
h
ha
s
b
e
e
n
s
h
ow
n
t
o
i
mp
r
ove
t
he
a
c
c
u
r
a
c
y
o
f
li
ne
a
r
c
las
s
i
f
ie
r
s
,
i
t
a
ls
o
e
xp
a
n
ds
t
he
s
pa
c
e
o
f
t
a
s
k
i
np
u
ts
,
wh
ic
h
c
a
n
r
e
s
u
lt
in
a
s
ig
ni
f
i
c
a
nt
in
c
r
e
a
s
e
in
tr
a
i
n
in
g
t
im
e
,
e
s
pe
c
i
a
l
ly
whe
n
de
a
li
n
g
wi
t
h
h
i
gh
-
d
i
men
s
i
ona
l
d
a
t
a
o
f
l
a
r
ge
vo
lu
me
.
Ho
we
v
e
r
,
in
t
his
pa
p
e
r
,
we
a
i
m
t
o
a
d
d
r
e
s
s
b
o
th
o
f
th
e
s
e
d
r
a
wb
a
c
ks
.
A
n
e
w
d
a
t
a
pa
r
t
i
ti
on
in
g
a
l
go
r
it
h
m
w
a
s
us
e
d
in
c
on
ju
nc
t
io
n
wi
th
a
dd
it
io
na
l
a
p
pl
ic
a
t
io
n
o
f
P
C
A
a
t
e
a
c
h
l
e
v
e
l
of
the
h
ie
r
a
r
c
h
ica
l
e
ns
e
mb
le
.
T
he
f
ir
s
t
a
p
p
r
o
a
c
h
s
ho
we
d
a
s
ig
ni
f
ica
nt
im
p
r
o
ve
men
t
i
n
t
he
a
c
c
u
r
a
c
y
o
f
t
he
e
n
s
e
m
b
le
m
e
th
od
,
wh
il
e
t
he
s
e
c
o
nd
a
p
p
r
oa
c
h
s
ig
n
if
ic
a
n
tl
y
r
e
du
c
e
d
i
ts
t
r
a
i
n
in
g
t
im
e
.
T
h
e
c
om
bi
ne
d
us
e
o
f
bo
th
a
pp
r
oa
c
he
s
p
r
ov
id
e
d
a
s
u
bs
t
a
n
ti
a
l
e
n
ha
n
c
e
men
t
in
th
e
pe
r
f
or
man
c
e
of
t
he
c
a
s
c
a
de
e
ns
e
mb
le
ba
s
e
d
o
n
t
wo
c
r
i
ti
c
a
l
i
nd
ic
a
t
o
r
s
.
L
e
t
us
take
a
c
los
e
r
look
a
t
thes
e
two
methods
.
I
n
t
his
pa
pe
r
,
we
p
r
opos
e
a
n
a
ddit
ional
us
e
of
P
C
A
to
r
e
duc
e
the
dim
e
ns
ionalit
y
o
f
the
input
da
ta
s
pa
c
e
f
o
r
e
a
c
h
we
a
k
c
las
s
if
ier
of
the
hier
a
r
c
hica
l
method.
T
he
ba
s
ic
c
a
s
c
a
de
e
ns
e
mb
le
[
33]
uti
li
s
e
s
a
qua
dr
a
ti
c
W
i
e
ne
r
polynom
ial
a
t
e
a
c
h
leve
l,
whic
h
c
a
n
incr
e
a
s
e
the
dim
e
ns
ionalit
y
of
the
input
da
ta
s
pa
c
e
.
T
his
a
ppr
oa
c
h
a
ll
ows
f
or
a
mo
r
e
e
f
f
icie
nt
a
nd
e
f
f
e
c
ti
ve
im
plem
e
ntation
of
the
method.
T
o
a
utom
a
te
thi
s
pr
oc
e
dur
e
,
we
us
e
d
the
metho
d
of
c
a
lcula
ti
ng
c
umul
a
ti
ve
va
r
ianc
e
e
xplaine
d
va
lues
.
I
t
ha
s
be
e
n
de
ter
mi
ne
d
that
the
number
of
pr
incipa
l
c
omponents
c
a
n
be
a
utom
a
ti
c
a
ll
y
s
e
lec
te
d
to
mee
t
the
us
e
r
's
s
pe
c
if
ied
pe
r
c
e
ntage
of
va
r
ianc
e
e
xplain
e
d.
R
e
c
e
nt
numer
ica
l
modelli
ng
r
e
s
ult
s
ha
ve
de
m
on
s
tr
a
ted
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
96
3
-
97
4
966
that
s
e
lec
ti
ng
a
va
lue
of
95%
gua
r
a
ntee
s
that
e
a
c
h
we
a
k
c
las
s
if
ier
c
ons
ider
s
the
f
unda
menta
l
inf
or
mation
ne
c
e
s
s
a
r
y
f
or
the
a
na
lys
is
.
Our
modi
f
ica
ti
on
s
ign
if
ica
ntl
y
r
e
duc
e
s
the
input
da
ta
s
pa
c
e
of
the
pr
o
blem
by
dis
c
a
r
ding
les
s
s
igni
f
ica
nt
or
nois
y
pr
incipa
l
c
omp
one
nts
.
T
his
r
e
duc
ti
on
is
by
mo
r
e
than
10
ti
mes
a
n
d
he
lps
to
s
hor
ten
the
tr
a
ini
ng
dur
a
ti
on
f
o
r
both
e
a
c
h
we
a
k
mac
hine
lea
r
ning
-
ba
s
e
d
c
las
s
if
ier
a
nd
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
a
s
a
whole
.
W
e
c
a
ll
the
us
e
of
thi
s
p
r
oc
e
dur
e
in
the
ba
s
e
li
ne
me
thod
[
33
]
modi
f
ica
ti
on
1
[
3
5]
.
B
oth
the
ba
s
ic
[
33]
a
nd
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
in
thi
s
pa
pe
r
r
e
quir
e
divi
ding
a
lar
ge
da
tas
e
t
int
o
pa
r
ts
to
f
or
m
a
c
a
s
c
a
de
s
tr
uc
tur
e
.
A
n
e
w
da
ta
pa
r
ti
ti
oning
a
lgor
it
hm
is
p
r
opos
e
d
in
the
pa
pe
r
f
o
r
f
or
mi
ng
s
ubs
e
ts
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
.
T
his
is
a
c
hieve
d
by
r
a
ndoml
y
s
e
lec
ti
ng
a
number
o
f
ve
c
tor
s
f
r
om
the
tr
a
ini
ng
s
e
t
that
c
or
r
e
s
ponds
to
the
us
e
r
-
de
f
ined
s
iz
e
.
T
hus
,
the
s
ubs
e
ts
f
or
the
im
p
r
ove
d
methods
c
a
n
be
e
it
he
r
lar
ge
r
o
r
s
maller
in
s
ize
c
ompar
e
d
to
the
s
ubs
e
ts
f
or
the
o
r
igi
na
l
method
[
33]
.
Additi
ona
ll
y,
the
s
ub
s
e
ts
may
c
ontain
r
e
pe
a
ted
ve
c
tor
s
,
whic
h
is
not
a
l
lowe
d
in
th
e
e
xis
ti
ng
method
[
33]
.
T
he
s
e
modi
f
ica
ti
ons
a
im
to
e
nha
nc
e
the
a
c
c
ur
a
c
y
of
e
a
c
h
we
a
k
c
las
s
if
ier
a
nd
the
im
pr
o
ve
d
S
GD
-
ba
s
e
d
c
a
s
c
a
d
e
e
ns
e
mbl
e
a
s
a
whole
.
T
he
a
ppr
oa
c
h
us
e
d
in
the
o
r
igi
na
l
method
[
33]
is
r
e
f
e
r
r
e
d
to
a
s
m
odif
ica
ti
on
2
[
36]
.
T
he
pe
r
f
o
r
manc
e
of
the
e
xis
ti
ng
c
a
s
c
a
de
e
ns
e
mbl
e
[
33]
c
a
n
be
e
nha
nc
e
d
by
c
ombi
ning
mo
dif
ica
ti
on
1
a
nd
modi
f
ica
ti
on
2.
How
e
ve
r
,
it
is
ne
c
e
s
s
a
r
y
to
modi
f
y
the
tr
a
ini
ng
a
lgor
it
hms
a
nd
a
pply
them
to
t
he
im
pr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
d
e
e
ns
e
mbl
e
.
P
lea
s
e
r
e
f
e
r
to
F
igur
e
1
f
or
the
f
low
c
ha
r
t
o
f
the
i
mpr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
tr
a
ini
ng.
L
e
t
us
e
xplor
e
the
ke
y
s
tage
s
of
the
e
nha
nc
e
d
tr
a
ini
ng
pr
oc
e
s
s
a
nd
the
a
ppli
c
a
ti
on
of
the
im
pr
ove
d
method
us
ing
modi
f
ica
ti
on
1
a
nd
modi
f
ica
ti
on
2
in
gr
e
a
ter
de
tail
.
I
n
or
de
r
to
do
s
o,
we
will
int
r
o
duc
e
the
c
onc
e
pt
of
da
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
,
whic
h
is
ut
il
ize
d
a
t
e
ve
r
y
leve
l
o
f
the
im
pr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
,
a
nd
outl
ine
it
s
ke
y
s
tage
s
.
T
he
da
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
c
ons
is
t
s
of
the
f
oll
owing
s
teps
:
i)
nor
maliza
ti
on
of
da
ta
by
c
olum
ns
ba
s
e
d
on
the
ma
xim
u
m
e
leme
nt;
ii
)
qua
dr
a
ti
c
e
xpa
ns
ion
of
the
no
r
malize
d
input
s
of
a
given
da
ta
s
a
mpl
e
via
the
wie
ne
r
polynom
ial;
a
nd
ii
i)
a
pply
ing
P
C
A
a
nd
s
e
lec
ti
ng
the
number
of
pr
incipa
l
c
omponents
that
pr
ovide
95%
of
the
e
xpl
a
ined
va
r
ianc
e
.
B
e
f
or
e
be
ginni
ng
the
tr
a
ini
ng
p
r
oc
e
dur
e
,
the
tr
a
ini
ng
da
tas
e
t
is
divi
de
d
int
o
s
ubs
e
ts
us
ing
a
ne
w
da
ta
pa
r
ti
ti
on
ing
a
lgor
i
thm
.
T
his
c
r
e
a
tes
N
-
s
ub
s
e
ts
with
r
e
pe
ti
ti
ons
,
whic
h
de
ter
m
ine
the
N
leve
ls
of
the
im
pr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
.
2.
1.
L
e
ar
n
in
g
a
lgori
t
h
m
−
S
tep
1.
T
he
da
ta
p
r
oc
e
s
s
ing
pr
oc
e
dur
e
is
pe
r
f
or
me
d
on
the
f
ir
s
t
s
ubs
e
t
to
tr
a
in
the
we
a
k
c
las
s
if
ier
1
(
S
GD
-
ba
s
e
d
c
las
s
if
ier
1)
.
−
S
t
e
p
2
.
T
h
e
d
a
ta
p
r
oc
e
s
s
i
ng
p
r
oc
e
du
r
e
is
pe
r
f
o
r
me
d
o
n
t
he
s
e
c
on
d
s
u
bs
e
t
a
n
d
a
p
pl
ie
d
to
th
e
p
r
e
-
t
r
a
i
ne
d
w
e
a
k
c
las
s
i
f
i
e
r
1
.
T
h
e
o
ut
pu
t
v
a
l
ue
s
o
bt
a
i
ne
d
f
r
om
S
G
D
-
b
a
s
e
d
c
l
a
s
s
i
f
ie
r
1
a
r
e
a
dde
d
t
o
s
u
bs
e
t
2
a
s
a
n
a
d
di
ti
o
na
l
f
e
a
t
ur
e
.
A
f
t
e
r
pe
r
f
o
r
mi
ng
th
e
da
ta
p
r
oc
e
s
s
in
g
p
r
oc
e
du
r
e
,
we
a
k
c
las
s
if
ie
r
2
(
S
G
D
-
ba
s
e
d
c
las
s
i
f
i
e
r
2)
is
t
r
a
in
e
d
.
−
S
tep
3.
T
he
d
a
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
is
pe
r
f
o
r
m
e
d
on
the
thi
r
d
s
ubs
e
t
a
nd
a
ppli
e
s
it
to
the
p
r
e
vious
ly
tr
a
ined
we
a
k
c
las
s
if
ier
1.
T
he
output
va
lues
obtaine
d
f
r
om
S
GD
-
ba
s
e
d
c
las
s
if
ier
1
a
r
e
a
dde
d
t
o
s
ubs
e
t
3
a
s
a
n
a
ddit
ional
f
e
a
tur
e
.
T
he
n,
pe
r
f
o
r
m
the
da
ta
pr
o
c
e
s
s
ing
pr
oc
e
dur
e
a
nd
a
pply
it
to
the
pr
e
-
tr
a
ined
we
a
k
c
l
a
s
s
i
f
i
e
r
2
.
T
h
e
o
u
t
p
u
t
v
a
l
u
e
s
o
b
t
a
i
n
e
d
f
r
o
m
S
G
D
-
b
a
s
e
d
c
l
a
s
s
i
f
i
e
r
2
a
r
e
a
d
d
e
d
t
o
s
u
b
s
e
t
3
a
s
a
n
a
d
d
i
t
i
o
n
a
l
f
e
a
t
u
r
e
.
F
i
n
a
l
l
y
,
a
f
t
e
r
p
e
r
f
o
r
m
i
n
g
t
h
e
d
a
t
a
p
r
o
c
e
s
s
i
n
g
p
r
o
c
e
d
u
r
e
,
w
e
a
k
c
l
a
s
s
i
f
i
e
r
3
(
S
G
D
-
b
a
s
e
d
c
l
a
s
s
i
f
i
e
r
3
)
i
s
t
r
a
i
n
e
d
.
−
S
tep
N.
T
he
da
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
is
pe
r
f
or
med
on
the
las
t,
s
ubs
e
t
N
a
nd
it
s
a
ppli
c
a
ti
on
to
the
pr
e
vious
ly
tr
a
ined
we
a
k
c
las
s
if
ier
1.
T
he
output
va
lues
obtaine
d
f
r
om
S
GD
-
ba
s
e
d
c
las
s
if
ier
1
a
r
e
a
dde
d
to
s
ubs
e
t
N
a
s
a
n
a
ddit
ional
f
e
a
tur
e
.
T
he
n,
pe
r
f
or
m
the
da
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
a
nd
a
pply
it
t
o
the
pr
e
-
tr
a
ined
we
a
k
c
las
s
if
ier
2.
T
he
output
va
lues
obtaine
d
f
r
om
S
GD
-
ba
s
e
d
c
las
s
if
ier
2
a
r
e
a
dde
d
to
s
ubs
e
t
N
a
s
a
n
a
ddit
ional
f
e
a
tur
e
.
T
he
n
,
pe
r
f
or
m
the
da
ta
p
r
oc
e
s
s
ing
pr
oc
e
dur
e
a
nd
a
pply
it
to
the
p
r
e
-
tr
a
ined
we
a
k
c
las
s
if
ier
3.
T
he
s
e
s
teps
a
r
e
r
e
pe
a
ted
a
t
e
a
c
h
s
ubs
e
que
nt
leve
l
unti
l
the
f
inal
leve
l
of
the
im
p
r
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
is
r
e
a
c
he
d,
whe
r
e
the
las
t
we
a
k
c
las
s
if
ier
(
S
GD
-
ba
s
e
d
c
las
s
if
ier
N)
is
t
r
a
i
ne
d.
2.
2.
App
li
c
a
t
ion
a
lgorit
h
m
I
n
the
a
ppli
c
a
ti
on
a
lgor
it
hm
f
or
the
im
p
r
ove
d
S
G
D
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
,
the
input
da
ta
ve
c
tor
i
s
c
las
s
if
ied
int
o
one
of
the
pr
oblem
-
de
f
ined
c
las
s
e
s
u
s
ing
a
pr
e
-
tr
a
ined
c
a
s
c
a
de
s
c
he
m
e
with
N
leve
ls
.
T
he
given
ve
c
tor
unde
r
goe
s
a
d
a
ta
pr
oc
e
s
s
ing
p
r
oc
e
dur
e
a
n
d
is
then
a
ppli
e
d
to
the
p
r
e
-
tr
a
ined
we
a
k
c
las
s
if
ier
1
.
T
he
output
va
lues
obtaine
d
f
r
om
S
GD
-
ba
s
e
d
c
las
s
if
ier
1
we
r
e
a
dde
d
to
the
given
ve
c
tor
a
s
a
n
a
ddit
ional
f
e
a
tur
e
.
T
he
n
,
the
d
a
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
wa
s
pe
r
f
or
med
a
nd
the
ve
c
tor
wa
s
a
ppli
e
d
to
the
pr
e
-
tr
a
ined
we
a
k
c
las
s
if
ier
2.
S
ubs
e
que
ntl
y,
the
output
va
lues
obtaine
d
f
r
om
S
GD
-
ba
s
e
d
c
las
s
if
ier
2
we
r
e
a
dde
d
to
t
he
given
da
ta
ve
c
tor
a
s
a
nother
a
ddit
ional
f
e
a
tur
e
.
T
he
d
a
ta
pr
oc
e
s
s
ing
pr
oc
e
dur
e
wa
s
pe
r
f
or
med
a
ga
in
a
nd
th
e
ve
c
tor
wa
s
a
ppli
e
d
to
the
pr
e
-
tr
a
ined
we
a
k
c
las
s
if
ier
3
.
All
the
s
teps
outl
ined
pr
e
vios
ly
a
r
e
c
a
r
r
ied
out
a
t
e
a
c
h
s
ubs
e
qu
e
nt
leve
l
unti
l
the
s
tate
leve
l
of
the
im
pr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
d
e
e
ns
e
mbl
e
is
r
e
a
c
he
d.
At
thi
s
po
int
,
the
f
inal
we
a
k
c
las
s
if
ier
(
S
GD
-
ba
s
e
d
c
las
s
if
ier
N)
is
a
ppli
e
d
to
de
ter
mi
ne
the
de
s
ir
e
d
membe
r
s
hip
c
las
s
of
the
input
da
ta
ve
c
tor
.
T
he
i
mpr
ove
d
c
a
s
c
a
de
e
n
s
e
mbl
e
of
f
e
r
s
the
f
o
ll
owing
a
dva
ntage
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
n
e
nhanc
e
d
c
as
c
ade
e
ns
e
mble
me
thod
for
big
da
ta
analys
is
(
I
v
an
I
z
onin
)
967
−
I
mpr
ove
s
the
ge
ne
r
a
li
z
a
ti
on
p
r
ope
r
ti
e
s
of
the
da
ta
c
las
s
if
ica
ti
on
method;
−
I
nc
r
e
a
s
e
s
the
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y;
−
R
e
duc
e
s
s
ub
s
a
mpl
e
dim
e
ns
ionalit
y
a
t
e
a
c
h
c
a
s
c
a
de
e
ns
e
mbl
e
leve
l;
−
R
e
duc
e
s
the
c
ompl
e
xit
y
of
c
omput
a
ti
on
o
f
the
s
e
lec
ted
we
a
k
c
las
s
if
ier
;
−
S
hor
tens
the
tr
a
ini
ng
pr
oc
e
dur
e
dur
a
ti
on.
F
igur
e
1.
F
low
-
c
ha
r
t
f
o
r
the
i
m
p
r
ov
e
d
S
G
D
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mb
le
(
t
r
a
i
n
in
g
m
od
e
)
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
T
o
s
im
ulate
the
ope
r
a
ti
on
of
the
im
p
r
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
,
the
a
uthor
s
c
r
e
a
ted
c
us
tom
s
of
twa
r
e
us
ing
the
P
ython
pr
ogr
a
mm
ing
langua
ge
,
ba
s
e
d
on
pr
inciples
f
r
om
[
37]
,
[
38
]
.
E
xpe
r
im
e
ntal
s
tud
ies
we
r
e
c
a
r
r
ied
out
on
a
c
omput
e
r
with
a
n
I
ntel®
C
or
e
™
i7
-
8750H
pr
oc
e
s
s
or
(
c
lock
f
r
e
que
nc
y
2.
20
GH
z
)
,
R
AM
8
GB
.
3.
1.
Dat
as
e
t
d
e
s
c
r
ip
t
ion
s
T
he
2021
United
S
tate
s
Dis
e
a
s
e
R
is
k
F
a
c
tor
S
ur
ve
il
lanc
e
S
ys
tem
(
B
R
F
S
S
)
pr
ovided
e
xtens
ive
da
tas
e
ts
,
whic
h
we
r
e
dis
s
e
mi
na
ted
by
the
C
e
nter
s
f
or
Dis
e
a
s
e
C
ontr
ol
a
nd
P
r
e
ve
nti
on
a
c
r
os
s
the
Unite
d
S
tate
s
a
nd
it
s
s
ur
ve
ye
d
r
e
gions
[
37]
.
T
he
2021
c
yc
le
of
th
e
B
R
F
S
S
inves
ti
ga
ted
a
r
a
nge
of
he
a
lt
h
pa
r
a
m
e
ter
s
,
s
uc
h
a
s
ove
r
a
ll
he
a
lt
h
a
s
s
e
s
s
ment,
dur
a
ti
on
of
we
ll
ne
s
s
,
phys
ica
l
a
c
ti
vit
y
leve
ls
,
hype
r
tens
ion
a
nd
c
h
oles
ter
ol
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
96
3
-
97
4
968
s
c
r
e
e
ning,
pr
e
va
lenc
e
of
c
hr
onic
il
lnes
s
a
nd
a
r
thr
i
ti
c
c
ondit
ions
,
tobac
c
o
us
a
ge
pa
tt
e
r
ns
,
f
r
uit
a
nd
v
e
ge
table
c
ons
umpt
ion
ha
bit
s
,
a
nd
a
c
c
e
s
s
ibi
li
ty
to
medic
a
l
a
s
s
is
tan
c
e
(
c
or
e
s
e
c
ti
on)
.
T
he
pr
im
a
r
y
da
tas
e
t
ini
ti
a
ll
y
c
ontaine
d
a
br
oa
d
s
pe
c
tr
um
of
inf
o
r
mation
r
e
late
d
to
thes
e
he
a
lt
h
c
ondit
ions
.
T
his
c
ompr
e
he
ns
ive
da
tas
e
t
pr
ovides
a
s
oli
d
f
ounda
ti
on
f
or
f
u
r
ther
a
na
lys
is
a
nd
r
e
s
e
a
r
c
h.
T
he
da
ta
s
e
t
wa
s
r
e
f
ined
to
f
oc
us
s
olely
on
li
f
e
s
tyl
e
f
a
c
tor
s
r
e
leva
nt
to
human
he
a
lt
h.
Ou
r
main
objec
ti
ve
wa
s
to
p
r
e
dict
oc
c
ur
r
e
nc
e
s
of
c
a
r
diovas
c
ular
dis
e
a
s
e
s
thr
ough
a
bina
r
y
c
las
s
if
ica
ti
on
tas
k.
T
he
r
e
s
ult
ing
da
tas
e
t
c
ons
is
t
s
of
308,
854
r
e
c
or
ds
,
e
a
c
h
with
29
a
t
tr
ibut
e
s
.
T
he
f
i
r
s
t
pha
s
e
o
f
da
ta
pr
e
pr
oc
e
s
s
ing
invol
ve
d
ide
nti
f
ying
a
nd
r
e
movi
ng
dupli
c
a
te
e
ntr
ies
to
e
ns
ur
e
the
uniquene
s
s
of
e
a
c
h
r
e
c
or
d.
T
his
s
tep
is
c
r
uc
ial
f
or
maintaining
the
int
e
gr
it
y
of
the
da
tas
e
t
a
nd
p
r
e
ve
nti
ng
r
e
dunda
nc
y
in
the
a
na
lys
is
.
B
y
e
li
mi
na
ti
ng
dupli
c
a
te
e
ntr
ies
,
the
r
e
s
e
a
r
c
he
r
s
a
im
e
d
to
s
tr
e
a
ml
ine
th
e
da
tas
e
t
a
nd
pr
e
pa
r
e
it
f
or
f
ur
ther
a
na
lys
is
.
Af
ter
e
ns
ur
ing
t
he
uniquene
s
s
of
e
a
c
h
r
e
c
or
d,
the
ne
xt
s
tep
wa
s
to
a
ddr
e
s
s
a
nd
r
e
c
ti
f
y
a
ny
mi
s
s
ing
va
lues
withi
n
the
da
tas
e
t
.
M
is
s
ing
va
lues
c
a
n
s
igni
f
ica
ntl
y
im
pa
c
t
the
qua
li
ty
a
nd
r
e
li
a
bil
it
y
of
the
a
na
lys
is
.
B
y
a
ddr
e
s
s
ing
a
nd
r
e
c
ti
f
ying
thes
e
mi
s
s
ing
va
lue
s
,
the
r
e
s
e
a
r
c
he
r
s
a
im
e
d
to
e
nha
nc
e
the
a
c
c
ur
a
c
y
a
nd
r
obus
tnes
s
of
the
da
tas
e
t
f
or
s
ubs
e
que
nt
a
na
lyt
ica
l
pr
oc
e
s
s
e
s
.
F
oll
owing
the
p
r
e
pr
oc
e
s
s
ing
s
tage
,
the
s
ubs
e
que
nt
a
na
lyt
ica
l
s
tep
f
oc
us
e
d
on
mi
ti
ga
ti
ng
the
c
las
s
im
ba
lanc
e
obs
e
r
ve
d
in
the
da
tas
e
t.
I
nit
ially,
the
dis
tr
ibut
ion
r
a
ti
o
be
twe
e
n
c
las
s
e
s
s
tood
a
t
92
%
to
8
%
.
M
a
int
a
ini
ng
ba
lanc
e
d
c
las
s
e
s
is
c
r
uc
ial
f
o
r
the
e
f
f
ica
c
y
of
mac
hine
lea
r
n
ing
models
,
a
s
it
c
a
n
s
igni
f
ica
ntl
y
im
pa
c
t
the
model's
a
bil
it
y
to
make
a
c
c
ur
a
te
pr
e
dictions
.
T
o
a
c
hieve
ba
lanc
e
d
c
las
s
e
s
,
two
pr
incipa
l
a
lg
or
it
hms
we
r
e
e
mpl
oye
d
c
onc
ur
r
e
ntl
y:
s
ynthetic
mi
nor
it
y
ove
r
s
a
mpl
ing
tec
hn
ique
(
S
M
OT
E
)
a
nd
Ne
a
r
M
is
s
.
S
M
OT
E
wa
s
us
e
d
to
a
ugment
ins
tanc
e
s
of
the
mi
nor
it
y
c
las
s
,
while
Ne
a
r
M
is
s
w
a
s
e
mpl
oye
d
to
r
e
duc
e
ins
tanc
e
s
of
the
major
it
y
c
las
s
.
T
his
it
e
r
a
ti
ve
pr
oc
e
s
s
invol
ve
d
a
djus
ti
ng
va
r
ious
pa
r
a
mete
r
va
lues
to
r
e
g
ulate
the
number
of
ins
tanc
e
s
f
r
om
both
c
las
s
e
s
,
a
i
mi
ng
to
a
c
hieve
a
mor
e
ba
lanc
e
d
dis
tr
ibut
ion
.
A
f
ter
a
thor
ough
it
e
r
a
ti
ve
p
r
oc
e
s
s
of
a
djus
ti
ng
pa
r
a
mete
r
va
lues
,
it
ha
s
be
e
n
de
ter
mi
ne
d
that
the
opti
mal
a
c
c
ur
a
c
y
a
nd
s
upe
r
ior
ge
ne
r
a
li
z
a
ti
on
f
o
r
the
us
e
d
c
las
s
if
ier
c
a
n
be
a
c
hieve
d
by
s
e
lec
ti
ng
e
xa
c
tl
y
75,
000
ins
tanc
e
s
f
r
om
e
a
c
h
c
l
a
s
s
in
the
or
igi
na
l
da
tas
e
t.
As
a
r
e
s
ult
,
the
f
inal
da
tas
e
t
f
or
im
pleme
nti
ng
mac
hine
lea
r
ning
tr
a
ini
ng
pr
oc
e
dur
e
s
c
ontains
150,
000
obs
e
r
va
ti
ons
.
3.
2.
Op
t
im
al
p
ar
am
e
t
e
r
s
s
e
lec
t
ion
f
or
t
h
e
im
p
r
ove
d
c
as
c
ad
e
e
n
s
e
m
b
le
T
he
s
e
lec
ti
on
of
p
a
r
a
mete
r
s
is
of
utm
os
t
im
por
tanc
e
f
or
the
im
p
r
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
n
s
e
mbl
e
.
T
his
e
ns
e
mbl
e
joi
ntl
y
e
mpl
oys
a
ne
w
da
ta
pa
r
ti
ti
oning
a
lgor
it
h
m
a
nd
a
ddit
ional
a
ppli
c
a
ti
on
of
P
C
A
a
t
e
a
c
h
leve
l
of
the
hier
a
r
c
hica
l
e
ns
e
mbl
e
.
I
t
is
c
r
uc
ial
to
d
e
ter
mi
ne
the
f
ol
l
owing:
−
T
he
opti
mal
number
o
f
leve
ls
f
o
r
the
c
a
s
c
a
de
e
ns
e
mbl
e
;
−
T
he
opti
mal
s
ize
(
%
of
the
tr
a
ini
ng
s
a
mpl
e
)
o
f
e
a
c
h
s
ubs
e
t
wa
s
r
a
ndoml
y
ge
ne
r
a
ted
with
r
e
pe
ti
ti
ons
of
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
a
c
c
or
ding
to
the
e
nha
nc
e
d
tr
a
ini
ng
da
ta
s
e
pa
r
a
ti
on
a
lgo
r
it
hm
;
−
T
he
opti
mal
number
o
f
pr
incipa
l
c
omponents
a
t
e
a
c
h
leve
l
of
the
e
ns
e
mbl
e
a
f
ter
a
pplyi
ng
P
C
A;
−
Optim
a
l
pa
r
a
mete
r
s
f
o
r
e
a
c
h
of
the
we
a
k
c
las
s
if
ier
s
.
T
he
opti
mal
pa
r
a
mete
r
s
of
S
GD
we
r
e
s
e
lec
ted
us
ing
the
gr
id
s
e
a
r
c
h
method
a
s
a
we
a
k
pr
e
dictor
a
t
e
a
c
h
leve
l
of
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
.
T
he
number
of
pr
incipa
l
c
omponents
us
e
d
in
the
hier
a
r
c
hica
l
method
wa
s
de
ter
mi
ne
d
ba
s
e
d
on
the
c
umul
a
ti
ve
v
a
r
ianc
e
e
xplaine
d.
T
his
a
ppr
oa
c
h
a
ll
ows
f
or
the
c
a
l
c
ulation
of
the
tot
a
l
va
r
iation
in
the
da
ta
e
xplaine
d
by
a
c
h
os
e
n
number
of
p
r
incipa
l
c
omponents
.
F
or
the
opti
mi
z
a
ti
on
of
the
im
pr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
,
we
s
e
lec
ted
the
pr
incipa
l
c
omponents
a
t
e
a
c
h
lev
e
l
of
the
c
a
s
c
a
de
that
a
c
c
ounted
f
or
95%
of
the
e
xplaine
d
va
r
ianc
e
.
T
he
s
e
c
omponents
we
r
e
then
us
e
d
a
s
input
s
f
or
e
a
c
h
we
a
k
c
las
s
if
ier
.
T
his
a
ppr
oa
c
h
e
ns
ur
e
d
that
the
r
e
quir
e
d
number
of
pr
incipa
l
c
omponents
f
o
r
e
a
c
h
we
a
k
pr
e
dictor
of
the
e
ns
e
mbl
e
wa
s
a
utom
a
ti
c
a
ll
y
s
e
lec
te
d,
r
e
s
ult
ing
in
opti
mi
z
e
d
pe
r
f
o
r
manc
e
.
T
he
im
p
lem
e
ntation
of
a
utom
a
ti
on
in
thi
s
pr
oc
e
du
r
e
s
igni
f
ica
ntl
y
r
e
duc
e
s
the
ti
me
r
e
quir
e
d
to
c
onduc
t
r
e
s
e
a
r
c
h
on
the
e
f
f
e
c
ti
ve
ne
s
s
of
the
method
a
nd
it
s
p
r
a
c
ti
c
a
l
a
ppli
c
a
ti
on.
I
n
thi
s
pa
pe
r
,
s
e
ve
r
a
l
e
xpe
r
im
e
ntal
s
tudi
e
s
we
r
e
c
o
nduc
ted
to
de
ter
m
ine
the
mos
t
e
f
f
icie
nt
va
lues
f
or
the
f
ir
s
t
two
pa
r
a
mete
r
s
in
or
de
r
to
opti
mi
z
e
the
e
f
f
e
c
ti
ve
ne
s
s
of
the
i
mpr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
.
T
his
a
r
ti
c
le
pr
e
s
e
nts
the
r
e
s
ult
s
o
f
e
xpe
r
im
e
nts
o
n
the
a
c
c
ur
a
c
y,
s
pe
e
d,
a
nd
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
of
the
pr
opos
e
d
hier
a
r
c
hica
l
method.
T
he
number
o
f
c
a
s
c
a
de
leve
ls
va
r
i
e
d
f
r
om
2
to
6,
a
nd
r
a
ndom
s
ubs
a
mpl
e
s
with
r
e
pe
ti
ti
ons
we
r
e
f
or
med
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
of
dif
f
e
r
e
nt
s
ize
s
(
r
a
nging
f
r
om
20%
to
90%
of
the
ini
ti
a
l
tr
a
ini
ng
s
a
mpl
e
with
a
s
tep
of
10%
)
.
T
he
r
e
s
ult
s
a
r
e
s
hown
in
F
igur
e
2
(
a
)
s
hows
the
r
e
s
ult
s
in
tr
a
ini
ng
mode,
a
nd
F
ig
ur
e
2
(
b
)
s
hows
the
r
e
s
ult
s
in
tes
t
mo
de
.
F
igur
e
2
de
mons
tr
a
tes
that
u
ti
li
z
ing
a
s
mall
pe
r
c
e
ntage
(
20%
-
30%
)
o
f
the
tr
a
ini
ng
s
a
mpl
e
to
c
r
e
a
te
r
a
ndom
s
ubs
a
mpl
e
s
with
r
e
pe
ti
ti
ons
yields
high
a
c
c
ur
a
c
y
dur
ing
the
tr
a
ini
ng
mode
o
f
the
c
a
s
c
a
de
e
ns
e
mbl
e
.
How
e
ve
r
,
thi
s
a
ppr
oa
c
h
may
ne
ga
ti
ve
ly
im
pa
c
t
it
s
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
.
C
onve
r
s
e
ly,
us
ing
r
a
ndom
s
ubs
a
mpl
e
s
with
a
volum
e
o
f
mor
e
than
50%
of
th
e
tr
a
ini
ng
s
a
mpl
e
e
nha
nc
e
s
a
c
c
ur
a
c
y
dur
ing
the
a
p
pli
c
a
ti
on
mode
but
may
r
e
s
ult
in
ove
r
tr
a
ini
n
g
of
the
method
.
All
of
thes
e
c
ha
r
a
c
ter
is
ti
c
s
a
pply
to
c
a
s
c
a
de
de
s
i
gns
with
2,
3
,
4
,
5
,
a
nd
6
leve
ls
.
I
t
is
im
por
tant
to
note
that
u
s
ing
lar
ge
s
ubs
a
mpl
e
s
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
may
incr
e
a
s
e
the
tr
a
ini
ng
ti
me
o
f
the
e
nti
r
e
method
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
n
e
nhanc
e
d
c
as
c
ade
e
ns
e
mble
me
thod
for
big
da
ta
analys
is
(
I
v
an
I
z
onin
)
969
(
a
)
(
b)
F
igur
e
2.
C
ha
nge
in
the
a
c
c
ur
a
c
y
of
the
im
pr
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
n
s
e
mbl
e
(
F
1
-
s
c
or
e
)
whe
n
c
ha
nging
the
%
o
f
t
r
a
ini
ng
s
a
mpl
e
us
a
ge
with
r
e
pe
ti
ti
on
a
nd
the
number
o
f
e
ns
e
mbl
e
leve
ls
:
(
a
)
f
or
tr
a
ini
ng
mo
de
a
nd
(
b)
f
or
tes
t
mode
F
igur
e
2
a
ls
o
il
lus
tr
a
tes
that
incr
e
a
s
ing
the
numbe
r
o
f
leve
ls
in
the
c
a
s
c
a
de
e
ns
e
mbl
e
(
be
yond
5)
f
or
pr
oc
e
s
s
ing
a
given
da
tas
e
t
is
not
r
e
c
omm
e
nde
d,
a
s
it
may
r
e
duc
e
the
ge
ne
r
a
li
z
a
ti
on
p
r
ope
r
ti
e
s
of
the
p
r
opos
e
d
de
s
ign.
W
e
s
ugge
s
t
that
the
opti
mal
pa
r
a
mete
r
s
f
or
s
olvi
ng
the
pr
o
blem
a
r
e
to
us
e
a
c
a
s
c
a
de
of
f
our
le
ve
ls
a
nd
to
f
or
m
r
a
ndom
s
ubs
a
mpl
e
s
a
t
e
a
c
h
leve
l
us
ing
40
%
of
the
t
r
a
ini
ng
s
a
mpl
e
.
T
he
s
e
r
e
s
ult
s
a
r
e
s
umm
a
r
ize
d
in
F
igur
e
3.
B
a
s
e
d
on
the
r
e
s
ult
s
pr
e
s
e
nt
in
F
ig
ur
e
3,
i
t
c
a
n
be
s
tate
d
that
the
S
GD
-
ba
s
e
d
c
a
s
c
a
d
e
e
ns
e
mb
le,
whe
n
tr
a
ined
on
s
ubs
a
mpl
e
s
c
ompr
is
ing
40%
of
the
tr
a
ini
ng
s
a
mpl
e
s
ize
,
e
xhibi
ts
the
highes
t
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
a
nd
a
c
c
ur
a
c
y
in
both
modes
.
I
t
is
wor
t
h
noti
ng
that
the
us
e
o
f
s
ubs
a
mpl
e
s
of
thi
s
s
ize
doe
s
not
s
igni
f
ica
ntl
y
incr
e
a
s
e
the
tr
a
ini
ng
ti
me
o
f
the
meth
od
c
ompar
e
d
to
the
us
e
of
lar
ge
r
s
ubs
a
mpl
e
s
.
F
igur
e
3.
T
he
be
s
t
pa
r
a
mete
r
s
f
o
r
the
i
m
p
r
o
ve
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mb
le
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
96
3
-
97
4
970
3.
3.
Re
s
u
lt
s
T
a
ble
1
pr
e
s
e
nts
the
r
e
s
ult
s
o
f
the
im
p
r
ove
d
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
us
ing
f
our
pe
r
f
or
manc
e
indi
c
a
tor
s
.
T
h
e
r
e
s
ul
ts
f
r
o
m
T
a
b
le
1
i
nc
lu
de
t
he
o
p
ti
mi
z
e
d
p
a
r
a
m
e
t
e
r
s
o
f
th
e
S
G
D
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mb
le
i
m
p
r
o
ve
d
in
t
his
a
r
t
ic
le
.
T
h
e
r
e
s
u
l
ts
c
l
e
a
r
ly
de
m
on
s
tr
a
te
it
s
s
up
e
r
i
or
pe
r
f
or
ma
nc
e
c
o
mp
a
r
e
d
to
ot
he
r
me
th
o
ds
.
T
a
ble
1.
R
e
s
ult
s
f
or
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
P
e
r
f
or
ma
nc
e
i
ndi
c
a
to
r
T
r
a
in
in
g mode
T
e
s
t
mode
P
r
e
c
is
io
n
0.806
0.884
R
e
c
a
ll
0.803
0.75
F1
-
s
c
or
e
0.802
0.801
T
r
a
in
in
g t
im
e
(
s
e
c
onds
)
0.31
-
3.
4.
Com
p
ar
is
on
an
d
d
is
c
u
s
s
ion
T
he
pr
opos
e
d
s
olut
ion,
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
,
wa
s
e
va
luate
d
f
or
it
s
e
f
f
e
c
ti
ve
ne
s
s
by
c
ompar
ing
it
wi
th
s
e
ve
r
a
l
s
im
il
a
r
methods
:
i)
m
e
t
hod
[
17]
(
S
GD
a
lgor
it
hm)
;
ii
)
m
e
thod
[
18]
(
e
xten
de
d
-
input
S
GD
)
;
ii
i
)
m
e
thod
[
33]
(
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
)
;
iv)
m
odif
ica
ti
on
1
(
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
with
P
C
A
on
e
a
c
h
c
a
s
c
a
de
leve
l)
[
35
]
;
a
nd
v
)
m
odif
ica
ti
on
2
(
b
a
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
with
s
ubs
a
mpl
e
s
with
r
e
pe
t
it
ions
on
e
a
c
h
c
a
s
c
a
de
leve
l)
[
36]
.
T
he
a
utho
r
s
thor
oughly
a
na
lyze
d
the
da
ta
s
e
t
s
tudi
e
d
in
thi
s
a
r
ti
c
le
a
nd
c
onf
idently
inves
ti
ga
ted
the
e
f
f
e
c
ti
ve
ne
s
s
of
e
a
c
h
method
ment
ioned
a
bove
.
T
he
y
e
xpe
r
tl
y
s
e
lec
ted
the
opti
mal
pa
r
a
mete
r
s
of
e
a
c
h
method
us
ing
the
gr
id
s
e
a
r
c
h
method.
T
he
r
e
s
ult
s
a
r
e
s
umm
a
r
ize
d
in
T
a
ble
2.
T
a
ble
2.
Opt
im
a
l
pa
r
a
mete
r
s
f
or
a
ll
inves
ti
ga
ted
m
e
thods
M
e
th
od
O
pt
im
a
l
pa
r
a
me
te
r
s
M
e
th
od [
17]
lo
s
s
=
'
lo
g'
, pe
na
lt
y=
’
l2
’
, a
lp
ha
=
0.0001.
M
e
th
od [
18]
S
G
D
c
la
s
s
if
ie
r
, qua
dr
a
ti
c
s
W
ie
ne
r
pol
ynomi
a
l.
M
e
th
od [
33]
C
a
s
c
a
d
e
e
ns
e
mbl
e
of
t
he
S
G
D
a
lg
or
it
hms
;
qua
dr
a
ti
c
s
W
ie
ne
r
pol
ynomi
a
l;
5 de
pt
h l
e
ve
ls
;
tr
a
in
in
g s
a
mpl
e
i
s
di
vi
de
d i
nt
o 5 e
qua
l
pa
r
ts
.
M
odi
f
ic
a
ti
on 1
[
35]
C
a
s
c
a
d
e
e
ns
e
mbl
e
of
t
he
S
G
D
a
lg
or
it
hms
;
qua
dr
a
ti
c
s
W
ie
ne
r
pol
ynomi
a
l;
6 de
pt
h l
e
ve
ls
;
in
put
da
ta
s
pa
c
e
i
s
r
e
duc
e
d i
n di
me
ns
io
n
a
li
ty
a
t
e
a
c
h l
e
ve
l
of
t
h
e
c
a
s
c
a
de
us
in
g P
C
A
,
e
ns
ur
in
g a
t
le
a
s
t
95%
of
t
he
va
r
ia
nc
e
.
M
odi
f
ic
a
ti
on 2
[
36]
C
a
s
c
a
d
e
e
ns
e
mbl
e
of
t
he
S
G
D
a
lg
or
it
hms
;
qua
dr
a
ti
c
s
W
ie
ne
r
pol
ynomi
a
l;
3 de
pt
h l
e
ve
ls
;
tr
a
in
in
g s
a
mpl
e
w
a
s
s
ubs
a
mpl
e
d r
a
ndoml
y a
t
e
a
c
h l
e
ve
l,
w
it
h 7
0%
of
t
he
s
a
mpl
e
be
in
g
s
e
le
c
t
e
d w
it
h r
e
pe
ti
ti
ons
.
P
r
opos
e
d s
ol
ut
io
n
C
a
s
c
a
d
e
e
ns
e
mbl
e
of
t
he
S
G
D
a
lg
or
it
hms
;
qua
dr
a
ti
c
s
W
ie
ne
r
pol
ynomi
a
l;
4 de
pt
h l
e
ve
ls
;
tr
a
in
in
g s
a
mpl
e
w
a
s
s
ubs
a
mpl
e
d r
a
ndoml
y a
t
e
a
c
h l
e
ve
l,
w
it
h 4
0%
of
t
he
s
a
mpl
e
be
in
g s
e
le
c
t
e
d w
it
h r
e
pe
ti
ti
ons
;
in
put
da
ta
s
p
a
c
e
i
s
r
e
duc
e
d i
n di
me
ns
io
n
a
li
ty
a
t
e
a
c
h l
e
ve
l
of
t
h
e
c
a
s
c
a
de
us
in
g P
C
A
,
e
ns
ur
in
g a
t
le
a
s
t
95%
of
t
he
va
r
ia
nc
e
.
T
wo
c
r
it
e
r
ia
we
r
e
s
e
lec
ted
to
c
ompar
e
the
e
f
f
e
c
ti
ve
ne
s
s
of
a
ll
the
methods
unde
r
s
tudy:
i)
F
1
-
s
c
or
e
in
tr
a
ini
ng
a
nd
tes
t
modes
;
a
nd
ii
)
tr
a
ini
ng
ti
me
(
in
s
e
c
onds
)
;
the
f
ir
s
t
c
r
it
e
r
ion
pr
ovides
a
n
oppor
tuni
ty
to
c
ompar
e
the
a
c
c
ur
a
c
y
of
a
ll
methods
in
a
ppli
c
a
ti
on
mode.
F
ur
ther
mor
e
,
the
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
of
e
a
c
h
method
c
a
n
be
e
va
luate
d
by
the
dif
f
e
r
e
nc
e
in
F
1
s
c
or
e
s
be
twe
e
n
tr
a
ini
ng
a
nd
tes
t
mode.
T
he
s
e
c
ond
c
r
it
e
r
ion
a
ll
ows
us
to
e
s
ti
mate
the
dur
a
ti
on
of
the
t
r
a
ini
ng
pr
oc
e
dur
e
f
or
the
s
e
lec
ted
method,
whic
h
is
c
r
uc
ial
whe
n
a
na
lyzing
lar
ge
da
tas
e
ts
.
F
igur
e
s
4
a
nd
5
s
how
the
c
ompar
is
on
r
e
s
ult
s
f
or
F
1
s
c
or
e
a
nd
tr
a
ini
ng
t
im
e
,
r
e
s
pe
c
ti
ve
ly,
ba
s
e
d
on
both
c
r
it
e
r
ia
.
A
f
ter
c
a
r
e
f
ul
a
na
lys
is
of
the
c
ompar
i
s
on
r
e
s
ult
s
pr
e
s
e
nted
in
F
igu
r
e
s
4
a
nd
5,
it
is
c
lea
r
th
a
t
the
method
[
17
]
,
whic
h
is
ba
s
e
d
on
the
high
-
s
pe
e
d
S
GD
a
lgor
it
hm
,
ha
s
the
f
a
s
tes
t
tr
a
ini
ng
ti
me
in
F
ig
ur
e
5,
but
e
xhibi
ts
the
lowe
s
t
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
a
s
s
hown
in
F
igu
r
e
4
.
How
e
ve
r
,
thi
s
method
s
ti
ll
de
mons
tr
a
tes
high
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
.
T
he
m
e
thod
[
18]
wa
s
a
bl
e
to
incr
e
a
s
e
the
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
of
the
da
ta
by
mor
e
than
5%
a
c
c
or
ding
to
the
F
1
-
s
c
or
e
in
F
igu
r
e
4,
th
r
o
ugh
the
c
ombi
na
ti
on
of
the
h
igh
-
s
pe
e
d
S
GD
a
lgor
i
thm
a
nd
the
qua
dr
a
ti
c
W
iene
r
polynom
ial.
How
e
ve
r
,
the
us
e
of
qua
dr
a
ti
c
W
ien
e
r
polynom
ial
s
igni
f
ica
ntl
y
incr
e
a
s
e
s
th
e
dim
e
ns
ionalit
y
of
the
p
r
oblem,
whic
h
in
tur
n
lea
ds
to
a
longer
tr
a
ini
ng
p
r
oc
e
dur
e
.
T
he
t
r
a
ini
ng
ti
me
f
or
thi
s
method
ha
s
incr
e
a
s
e
d
s
igni
f
ica
ntl
y
c
ompar
e
d
t
o
the
p
r
e
vious
method
s
e
e
F
igur
e
5.
F
u
r
ther
mor
e
,
the
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
n
e
nhanc
e
d
c
as
c
ade
e
ns
e
mble
me
thod
for
big
da
ta
analys
is
(
I
v
an
I
z
onin
)
971
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
of
thi
s
method
ha
ve
de
ter
i
or
a
ted.
S
pe
c
if
ica
ll
y,
the
di
f
f
e
r
e
nc
e
be
twe
e
n
the
F
1
-
s
c
or
e
in
both
tr
a
ini
ng
modes
is
s
igni
f
ica
ntl
y
highe
r
than
that
of
method
[
17
]
.
T
he
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
(
method
[
33]
)
ha
s
be
e
n
s
hown
to
a
c
hieve
a
10%
higher
a
c
c
ur
a
c
y
(
F
1
-
s
c
or
e
)
c
ompar
e
d
to
m
e
thod
[
17]
a
nd
a
lm
os
t
5
%
c
ompar
e
d
to
method
[
18]
.
F
u
r
ther
mor
e
,
thi
s
me
thod
ha
s
the
a
dva
ntage
of
r
e
duc
ing
the
dur
a
t
ion
of
the
tr
a
ini
ng
pr
oc
e
dur
e
by
a
lm
os
t
ha
lf
c
ompar
e
d
to
method
[
18]
.
B
ut
the
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
ha
s
the
wor
s
t
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
of
a
ll
the
methods
inves
ti
ga
ted
a
s
s
hown
in
F
igur
e
4.
T
he
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
(
method
[
33]
)
wa
s
opti
mi
z
e
d
by
i
mpl
e
menting
P
C
A
a
t
e
a
c
h
leve
l
(
m
odif
ica
ti
on
1)
,
r
e
s
ult
ing
in
a
r
e
duc
ti
on
of
tr
a
in
in
g
ti
me
by
a
lm
os
t
7
ti
mes
.
F
u
r
ther
mor
e
,
m
odif
ica
ti
on
1
led
to
a
2
%
incr
e
a
s
e
in
a
c
c
ur
a
c
y
in
the
a
ppli
c
a
ti
on
mo
de
.
T
he
s
e
be
ne
f
it
s
a
r
e
a
tt
r
ibut
e
d
to
the
s
ubs
tantial
r
e
duc
ti
on
in
s
ubs
a
mpl
e
dim
e
ns
ionalit
y
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
.
T
he
number
of
pr
oblema
ti
c
inp
uts
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
wa
s
e
f
f
e
c
ti
ve
ly
r
e
duc
e
d
by
mor
e
than
tenf
old
by
us
ing
P
C
A,
whic
h
a
c
c
ounts
f
or
95%
of
the
va
r
ianc
e
.
T
he
s
e
lea
s
t
s
igni
f
ica
nt
p
r
inc
ipal
c
omponents
e
it
he
r
do
not
inf
luenc
e
the
c
las
s
if
ica
ti
on
r
e
s
ult
s
or
a
r
e
nois
e
c
ompon
e
nts
ne
ga
ti
ve
ly
a
f
f
e
c
ti
ng
the
c
las
s
if
ica
ti
on
r
e
s
ult
s
.
T
he
a
c
c
ur
a
c
y
of
the
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
(
method
[
33]
)
wa
s
im
pr
ove
d
by
im
ple
menting
a
ne
w
s
ubs
a
mpl
ing
s
c
he
me
f
or
e
a
c
h
c
a
s
c
a
de
leve
l
(
m
odif
ica
ti
on
2
)
.
T
his
r
e
s
ult
e
d
in
a
2%
incr
e
a
s
e
in
a
c
c
ur
a
c
y
ba
s
e
d
o
n
F
1
-
s
c
or
e
a
s
s
hown
in
F
igur
e
4
.
Additi
ona
ll
y,
th
is
modi
f
ica
ti
on
s
igni
f
ica
ntl
y
e
n
ha
nc
e
d
the
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
c
ompar
e
d
to
the
m
e
thod
[
33
]
.
None
thele
s
s
,
the
incr
e
a
s
e
d
s
ubs
a
mpl
e
s
ize
a
t
e
a
c
h
leve
l
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
led
to
a
tr
a
ini
ng
pr
oc
e
dur
e
that
wa
s
ne
a
r
ly
twice
a
s
long,
a
s
il
lus
tr
a
t
e
d
in
F
igu
r
e
5.
F
igur
e
4.
F
1
-
s
c
or
e
f
or
a
ll
inves
ti
ga
ted
methods
F
igur
e
5.
T
r
a
ini
ng
t
im
e
(
in
s
e
c
onds
)
f
or
a
ll
inves
ti
ga
ted
methods
T
he
c
ombi
ne
d
us
e
of
the
ne
w
s
ubs
a
mpl
ing
s
c
he
me
a
nd
the
a
ppli
c
a
ti
on
of
P
C
A
a
t
e
a
c
h
leve
l
o
f
the
ba
s
ic
c
a
s
c
a
de
e
ns
e
mbl
e
(
method
[
33
]
)
,
whic
h
is
pr
opos
e
d
in
thi
s
a
r
ti
c
le
(
pr
opos
e
d
s
olut
ion)
,
pr
ovi
de
d
high
a
c
c
ur
a
c
y,
the
highes
t
ge
ne
r
a
li
z
a
ti
on,
a
nd
s
igni
f
ic
a
ntl
y
lowe
r
tr
a
ini
ng
t
im
e
c
ompar
e
d
to
the
b
a
s
ic
c
a
s
c
a
d
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14
,
No.
2
,
Apr
il
20
25
:
96
3
-
97
4
972
e
ns
e
mbl
e
a
s
s
hown
in
F
igur
e
s
4
a
nd
5.
T
he
s
e
a
dva
ntage
s
make
thi
s
method
a
ve
r
y
pr
a
c
ti
c
a
l
s
olut
io
n.
F
utur
e
r
e
s
e
a
r
c
h
c
ould
be
dir
e
c
ted
towa
r
ds
two
main
a
r
e
a
s
:
‒
T
he
a
c
c
ur
a
c
y
of
the
im
pr
ove
d
c
a
s
c
a
de
e
ns
e
mbl
e
c
a
n
be
e
nha
nc
e
d
by
us
ing
a
lt
e
r
na
ti
ve
l
inea
r
metho
ds
a
s
we
a
k
c
las
s
if
ier
s
.
I
t
is
im
por
tant
to
c
ons
ider
thi
s
op
ti
on
only
i
f
thes
e
methods
ha
ve
be
e
n
pr
ove
n
to
pr
ovide
higher
a
c
c
ur
a
c
y
than
S
GD
whe
n
a
na
lyzing
a
s
pe
c
if
ic
da
tas
e
t.
‒
I
mpl
e
menting
a
lt
e
r
na
t
ive
pr
incipa
l
c
omponent
e
xt
r
a
c
ti
on
methods
,
s
uc
h
a
s
ne
ur
a
l
ne
twor
k
a
na
logu
e
s
of
P
C
A,
c
ould
potentially
r
e
duc
e
the
du
r
a
ti
on
o
f
th
e
tr
a
ini
ng
pr
oc
e
dur
e
f
o
r
the
e
nti
r
e
c
a
s
c
a
de
e
ns
e
mbl
e
,
pr
ovided
that
they
ha
ve
a
f
a
s
ter
pe
r
f
or
manc
e
than
t
he
c
las
s
ica
l
P
C
A.
How
e
ve
r
,
both
o
f
t
he
a
bove
a
ppr
oa
c
he
s
s
hould
be
us
e
d
taking
int
o
a
c
c
ount
the
s
pe
c
if
ics
of
a
pa
r
ti
c
ular
tas
k
a
nd
the
qua
li
ty,
qua
nti
ty
a
nd
dim
e
ns
ionalit
y
o
f
the
tr
a
ini
ng
da
ta
s
e
t
a
va
il
a
ble
to
s
olve
it
with
mac
hine
l
e
a
r
ning.
4.
CONC
L
USI
ON
T
he
incr
e
a
s
e
in
digi
tal
da
ta
pr
e
s
e
nts
both
c
ha
ll
e
nge
s
a
nd
oppor
tuni
ti
e
s
.
How
e
ve
r
,
with
the
va
s
t
volum
e
a
nd
c
ompl
e
xit
y
of
big
da
ta
,
tr
a
dit
ional
pr
oc
e
s
s
ing
methods
c
a
n
be
s
tr
a
ined
.
M
a
c
hine
lea
r
ning
is
a
ke
y
s
olut
ion
that
e
na
bles
a
na
lys
is
a
nd
ins
ight
e
xtr
a
c
ti
on
f
r
om
l
a
r
ge
da
tas
e
ts
.
T
he
int
e
gr
a
ti
on
of
mac
hine
lea
r
ning
a
nd
big
da
ta
ha
s
e
volved
s
igni
f
ica
ntl
y,
with
c
a
s
c
a
de
e
ns
e
mbl
e
s
,
pa
r
ti
c
ular
ly
e
ns
e
mbl
e
methods
,
s
howing
pr
omi
s
e
.
W
he
n
de
s
igni
ng
c
a
s
c
a
de
e
n
s
e
mbl
e
s
,
it
is
c
r
uc
ial
t
o
ba
lanc
e
the
f
a
c
tor
s
of
high
a
c
c
ur
a
c
y
a
nd
lengthy
tr
a
ini
ng,
e
s
pe
c
ially
with
c
ompl
e
x
da
tas
e
ts
a
nd
nonl
inea
r
tec
hniques
.
How
e
ve
r
,
with
c
a
r
e
f
ul
c
ons
ider
a
ti
on
a
nd
e
xpe
r
ti
s
e
,
c
a
s
c
a
de
e
ns
e
mbl
e
s
c
a
n
s
ti
ll
a
c
hieve
im
pr
e
s
s
ive
a
c
c
ur
a
c
y.
Additi
ona
ll
y
,
pa
r
ti
ti
on
ing
da
tas
e
ts
int
o
s
u
bs
e
ts
c
a
n
li
mi
t
mac
h
ine
lea
r
ning
a
lgo
r
it
hms
'
a
c
c
e
s
s
to
the
e
nt
ir
e
da
tas
e
t,
potentially
a
f
f
e
c
ti
ng
the
ove
r
a
ll
pe
r
f
or
manc
e
of
the
c
a
s
c
a
de
model.
T
his
pa
pe
r
pr
e
s
e
nts
s
igni
f
ica
nt
im
pr
ove
ments
to
the
S
GD
-
ba
s
e
d
c
a
s
c
a
de
e
ns
e
mbl
e
by
int
e
gr
a
ti
ng
a
ne
w
tr
a
ini
ng
da
ta
pa
r
ti
ti
oning
a
lgo
r
i
thm
a
nd
P
C
A
a
t
e
a
c
h
leve
l.
T
he
c
ombi
ne
d
us
e
of
thes
e
methods
e
nha
nc
e
s
the
e
ns
e
mbl
e
'
s
a
c
c
ur
a
c
y
a
nd
r
e
duc
e
s
tr
a
ini
ng
ti
me.
T
he
pa
pe
r
de
mons
tr
a
tes
thr
ough
modeling
that
the
c
a
s
c
a
de
e
ns
e
mbl
e
's
pe
r
f
or
manc
e
metr
ics
,
including
a
c
c
ur
a
c
y,
t
r
a
ini
ng
t
im
e
,
a
nd
ge
ne
r
a
l
iza
ti
on
pr
ope
r
ti
e
s
,
a
r
e
s
igni
f
ica
ntl
y
im
pr
ov
e
d
c
ompar
e
d
to
the
ba
s
e
li
ne
method.
F
utu
r
e
r
e
s
e
a
r
c
h
will
e
xplor
e
a
lt
e
r
na
ti
ve
methodologi
e
s
,
s
uc
h
a
s
non
-
it
e
r
a
ti
ve
a
r
ti
f
icia
l
ne
ur
a
l
ne
two
r
ks
(
s
uc
c
e
s
s
ive
g
e
ometr
ic
tr
a
ns
f
or
mations
model
(
S
G
T
M
)
ne
ur
a
l
-
li
ke
s
tr
uc
tu
r
e
)
a
nd
ne
ur
a
l
ne
twor
k
-
ba
s
e
d
va
r
iations
of
P
C
A,
to
e
nha
nc
e
a
c
c
ur
a
c
y,
r
e
duc
e
tr
a
ini
ng
ti
me
,
a
nd
maintain
ge
ne
r
a
li
z
a
ti
on
pr
ope
r
ti
e
s
with
c
onf
idenc
e
.
F
u
r
ther
mor
e
,
the
e
xa
mi
na
ti
on
of
the
c
a
s
c
a
de
e
ns
e
mbl
e
's
pr
e
s
e
nt
a
ti
on
a
s
a
polynom
ial
s
c
he
me
(
whe
n
uti
li
z
ing
S
GT
M
ne
ur
a
l
-
li
ke
s
tr
uc
tur
e
a
s
a
we
a
k
c
la
s
s
if
ier
f
or
the
c
a
s
c
a
de
)
is
int
e
nde
d
to
a
c
c
e
ler
a
te
inf
e
r
e
nc
e
ti
me
dur
ing
a
ppli
c
a
ti
on
s
tage
s
.
T
he
s
e
inqui
r
ies
ha
ve
the
potential
to
im
pr
ove
t
he
c
a
pa
bil
it
ies
of
c
a
s
c
a
de
e
n
s
e
mbl
e
s
a
nd
br
oa
de
n
their
a
ppli
c
a
bil
it
y
in
r
e
a
l
-
wo
r
ld
big
da
ta
s
c
e
na
r
ios
.
AC
KNOWL
E
DGE
M
E
NT
S
P
r
of
.
M
icha
l
Gr
e
gus
wa
s
s
uppor
ted
by
the
S
lova
k
R
e
s
e
a
r
c
h
a
nd
De
ve
lopm
e
nt
Age
nc
y
unde
r
the
c
ontr
a
c
t
No.
APVV
19
-
0581
.
T
his
wo
r
k
is
f
unde
d
by
the
E
ur
ope
a
n
Union’
s
Ho
r
izon
E
ur
ope
r
e
s
e
a
r
c
h
a
nd
innovation
pr
ogr
a
m
unde
r
gr
a
nt
a
gr
e
e
ment
No
10
1138678,
p
r
ojec
t
Z
E
B
AI
(
I
nnova
ti
ve
methodologi
e
s
f
or
the
de
s
ign
of
Z
e
r
o
-
E
mi
s
s
ion
a
nd
c
os
t
-
e
f
f
e
c
ti
ve
B
uil
di
ngs
e
nha
nc
e
d
by
Ar
ti
f
icia
l
I
ntelli
ge
nc
e
)
.
RE
F
E
RE
NC
E
S
[
1]
A
.
C
ha
ha
l,
P
. G
ul
ia
,
a
nd
N
.
S
. G
i
ll
,
“
D
if
f
e
r
e
nt
a
na
ly
ti
c
a
l
f
r
a
me
w
or
ks
a
nd
bi
gda
ta
mode
l
f
o
r
in
te
r
ne
t
of
th
in
gs
,”
I
ndone
s
ia
n
J
ou
r
nal
of
E
le
c
tr
ic
al
E
ngi
ne
e
r
in
g and C
om
put
e
r
S
c
ie
nc
e
, vol
. 25, no. 2,
pp.
1159
–
1166, 2022, doi:
10.11591/i
je
e
c
s
.v25.i2.pp1159
-
1166.
[
2]
I
. K
r
a
k, O
. S
te
li
a
, A
. P
a
s
hko, M
. E
f
r
e
mov, a
nd O
. K
hor
oz
ov, “
E
le
c
tr
oc
a
r
di
ogr
a
m c
la
s
s
if
ic
a
ti
on us
in
g w
a
ve
le
t
tr
a
ns
f
or
ma
ti
ons
,”
in
P
r
oc
e
e
di
ngs
-
15t
h
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
A
dv
anc
e
d
T
r
e
nds
in
R
adi
oe
le
c
tr
oni
c
s
,
T
e
le
c
o
m
m
uni
c
at
io
ns
and
C
om
p
ut
e
r
E
ngi
ne
e
r
in
g, T
C
SE
T
2020
, 2020, pp. 930
–
933
, doi
:
10.1109/T
C
S
E
T
49122.2020.235573.
[
3]
L
.
M
oc
hur
a
d
a
nd
N
.
K
r
yvi
ns
ka
,
“
P
a
r
a
ll
e
li
z
a
ti
on
of
f
in
di
ng
th
e
c
ur
r
e
nt
c
oor
di
na
te
s
of
th
e
li
da
r
ba
s
e
d
on
th
e
ge
ne
ti
c
a
lg
o
r
it
hm
a
nd
ope
nmp t
e
c
hnol
ogy,”
Sy
m
m
e
t
r
y
, vol
. 13, no. 4, 2021, doi
:
10.3
390/
s
ym13040666.
[
4]
A
.
C
ha
udha
r
y,
K
.
R
.
B
a
twa
da
,
N
.
M
it
ta
l,
a
nd
E
.
S
.
P
il
li
,
“
A
dM
a
p:
a
f
r
a
me
w
or
k
f
or
a
dve
r
ti
s
in
g
us
in
g
M
a
pR
e
duc
e
pi
pe
li
ne
,”
C
om
put
e
r
Sc
ie
n
c
e
and I
nf
or
m
at
io
n T
e
c
hnol
ogi
e
s
, vol
. 3, no. 2,
pp. 82
–
93, 2022, doi:
10.11591/cs
it
.v3i
2.pp82
-
93.
[
5]
A
.
H
.
A
l
-
H
a
ma
mi
a
nd
A
.
A
.
F
la
yyi
h,
“
E
nha
nc
in
g
bi
g
d
a
ta
a
na
ly
s
is
by
u
s
in
g
ma
p
-
r
e
duc
e
te
c
hni
qu
e
,”
B
ul
le
ti
n
of
E
le
c
tr
i
c
al
E
ngi
ne
e
r
in
g and I
nf
or
m
at
ic
s
, vol
. 7, no. 1, pp. 113
–
116, 2018,
do
i:
10.11591/ee
i.
v7i
1.895.
[
6]
Y
.
M
a
r
z
ha
n,
K
.
T
a
l
s
hyn,
K
.
K
a
ir
a
t,
B
.
S
a
ul
e
,
A
.
K
a
r
ly
ga
s
h,
a
nd
O
.
Y
e
r
bol
,
“
S
ma
r
t
te
c
hnol
ogi
e
s
of
th
e
r
is
k
-
ma
na
ge
me
nt
a
nd
de
c
is
io
n
-
ma
ki
ng
s
ys
t
e
ms
in
a
f
uz
z
y
da
ta
e
nvi
r
onme
nt
,”
I
ndo
ne
s
ia
n
J
our
nal
of
E
le
c
t
r
ic
al
E
ngi
ne
e
r
in
g
and
C
o
m
put
e
r
Sc
ie
nc
e
,
vol
. 28, no. 3, pp. 1463
–
1474, 2022, doi:
10.11591/i
je
e
c
s
.v28.i3.pp1463
-
1474.
[
7]
C
.
G
a
ngul
i,
S
.
K
.
S
ha
ndi
ly
a
,
M
.
N
e
hr
e
y,
a
nd
M
.
H
a
vr
yl
iu
k,
“
A
da
pt
iv
e
a
r
ti
f
ic
ia
l
be
e
c
ol
ony
a
lg
or
it
h
m
f
o
r
na
tu
r
e
-
in
s
pi
r
e
d
c
ybe
r
de
f
e
ns
e
,”
Sy
s
te
m
s
, vol
. 11, no.
1, 2023, doi:
10.3390/s
y
s
te
ms
11
010027.
[
8]
L
.
M
oc
hur
a
d,
K
.
S
ha
khovs
ka
,
a
nd
S
.
M
ont
e
ne
gr
o,
“
P
a
r
a
ll
e
l
s
o
lv
in
g
of
f
r
e
dhol
m
in
te
gr
a
l
e
qua
ti
ons
of
th
e
f
ir
s
t
ki
nd
by
T
ik
ho
nov
r
e
gul
a
r
iz
a
ti
on
me
th
od
us
in
g
O
pe
n
M
P
te
c
hnol
ogy,”
A
dv
anc
e
s
in
I
nt
e
ll
ig
e
nt
Sy
s
t
e
m
s
and
C
o
m
put
in
g
,
pp.
25
–
35,
2020,
doi
:
10.1007/978
-
3
-
030
-
33695
-
0_3.
[
9]
O
.
B
is
ik
a
lo
,
V
.
K
ha
r
c
he
nko,
V
.
K
ovt
un,
I
.
K
r
a
k,
a
nd
S
.
P
a
vl
ov,
“
P
a
r
a
me
te
r
iz
a
ti
on
of
th
e
s
to
c
ha
s
ti
c
mode
l
f
or
e
va
lu
a
ti
ng
va
r
ia
bl
e
s
ma
ll
da
ta
i
n t
he
S
ha
nnon e
nt
r
opy ba
s
i
s
,”
E
nt
r
opy
, vol
. 25, no.
2, 2023, doi:
10.3390/e25020184
.
Evaluation Warning : The document was created with Spire.PDF for Python.