T
E
L
KO
M
NI
K
A
,
V
ol
.
14,
N
o.
3,
S
ept
em
ber
20
16,
pp.
10
52
~
105
8
I
S
S
N
:
1
693
-
6
930
,
ac
c
r
edi
t
ed
A
b
y
D
IK
T
I,
D
e
c
r
e
e
N
o
:
58/
D
I
K
T
I
/
K
ep/
2013
D
O
I
:
10.
12928/
T
E
LK
O
M
N
I
K
A
.
v
1
4
i
3
.
3150
10
52
R
ec
ei
v
ed
N
ov
e
mber
25
,
20
1
5
;
R
ev
i
s
ed
J
u
ne
1
4
,
2
01
6
;
A
c
c
ept
e
d
J
un
e
2
9
,
20
1
6
A
C
o
mpa
r
i
s
on
of R
e
t
w
e
e
t P
r
e
di
c
ti
o
n
A
p
p
r
oa
c
h
e
s
: T
he
S
upe
r
i
or
i
t
y
o
f R
a
nd
om
For
e
s
t L
e
a
r
ni
n
g M
e
t
ho
d
H
en
d
r
a B
u
n
yam
i
n
*
1
,
T
o
m
a
s
T
u
n
y
s
2
1
D
epar
t
m
ent
of
I
nf
or
m
at
i
c
s
,
M
ar
anat
ha C
hr
i
s
t
i
an U
n
i
v
er
s
i
t
y
J
l
.
P
r
o
f
.
dr
g.
S
ur
i
a S
u
m
ant
r
i
N
o.
65 B
andu
ng
,
I
ndo
nes
i
a,
T
el
p
/
F
a
x
:
+6
2
-
222
0121
86/
2
220
05
915
2
D
epar
t
m
ent
of
C
o
m
put
er
S
c
i
e
nc
e
,
C
z
ec
h T
ec
hn
i
c
a
l
U
ni
v
er
s
i
t
y
Z
i
k
ov
a 1903
/
4
1
66 3
6 P
r
ag
ue
6,
C
z
ec
h R
e
publ
i
c
,
T
el
p:
+
420
-
22435
757
6
*
C
or
r
es
po
ndi
ng a
ut
ho
r
,
e
-
ma
i
l
:
hendr
a.
bu
ny
am
i
n@
i
t
.
m
ar
an
at
ha.
ed
u
1
, t
u
n
y
s
to
m
@
fe
l
.c
v
u
t.
c
z
2
A
b
st
r
act
W
e
c
o
ns
i
der
t
he
f
ol
l
ow
i
n
g r
et
w
eet
pr
edi
c
t
i
on t
as
k
:
g
i
v
e
n
a t
w
eet
,
pr
e
di
c
t
w
het
her
i
t
w
i
l
l
b
e
r
et
w
eet
e
d.
I
n
t
he
p
as
t
,
a
w
i
de
r
ange
of
l
ear
ni
n
g
m
et
h
ods
an
d
f
eat
ur
es
ha
s
been
pr
o
po
s
e
d
f
o
r
th
i
s
ta
s
k
. W
e
pr
ov
i
de
a
s
y
s
t
em
at
i
c
c
om
par
i
s
on
of
t
he
per
f
or
m
anc
e
of
t
h
e
s
e
l
e
ar
ni
n
g
m
et
hod
s
an
d
f
eat
ur
es
i
n
t
er
m
s
of
pr
edi
c
t
i
o
n a
c
c
ur
ac
y
an
d f
ea
t
ur
e i
m
por
t
a
nc
e.
S
pe
c
i
f
i
c
a
l
l
y
,
f
r
o
m
eac
h pr
ev
i
ou
s
l
y
pub
l
i
s
hed
a
ppr
oa
c
h w
e t
ak
e
t
he be
s
t
per
f
or
m
i
n
g f
e
at
ur
e
s
a
nd gr
o
up t
hes
e i
n
t
o t
w
o
s
et
s
:
us
er
f
eat
ur
es
and
t
w
eet
f
eat
ur
es
.
I
n ad
di
t
i
on
,
w
e c
ont
r
a
s
t
f
i
v
e l
ea
r
ni
n
g m
et
h
ods
,
bo
t
h l
i
near
an
d non
-
l
i
ne
ar
.
O
n t
op of
t
hat
,
w
e ex
am
i
ne t
he adde
d v
al
ue
of
a
pr
ev
i
ou
s
l
y
pr
opo
s
ed
t
i
m
e
-
s
en
s
i
t
i
v
e
m
od
el
i
ng
appr
oac
h.
T
o
t
he
aut
ho
r
s
’
k
now
l
edge
t
hi
s
i
s
t
he
f
i
r
s
t
at
t
em
pt
t
o c
ol
l
e
c
t
bes
t
p
er
f
o
r
m
i
ng f
eat
ur
e
s
and c
ont
r
a
s
t
l
i
ne
ar
and non
-
l
i
n
ear
l
ear
ni
n
g m
et
hods
.
W
e
per
f
or
m
our
c
om
p
ar
i
s
ons
on
a s
i
ngl
e da
t
as
et
a
nd f
i
nd
t
hat
us
er
f
eat
ur
e
s
s
uc
h
as
t
he
nu
m
ber
of
t
i
m
es
a
us
er
i
s
l
i
s
t
e
d,
n
um
ber
o
f
f
o
l
l
ow
er
s
,
and
av
er
ag
e n
um
ber
of
t
w
e
et
s
pu
bl
i
s
he
d p
er
da
y
m
os
t
s
t
r
ong
l
y
c
ont
r
i
but
e t
o pr
edi
c
t
i
on a
c
c
ur
ac
y
ac
r
o
s
s
s
e
l
ec
t
ed
l
ear
n
i
ng
m
et
hods
.
W
e a
l
s
o
f
i
n
d t
ha
t
a
r
andom
f
or
es
t
-
bas
e
d
l
e
ar
ni
ng,
w
h
i
c
h
has
n
ot
bee
n
em
pl
o
y
e
d
i
n
pr
ev
i
ou
s
s
t
udi
e
s
,
a
c
h
i
ev
es
t
he
hi
g
he
s
t
pe
r
f
o
r
m
a
nc
e
am
ong t
he
l
ear
n
i
ng m
et
h
od
s
w
e c
on
s
i
de
r
.
W
e a
l
s
o f
i
nd t
hat
on t
op of
pr
op
er
l
y
t
une
d l
ear
n
i
ng m
et
ho
ds
t
h
e
benef
i
t
s
of
t
i
m
e
-
s
en
s
i
t
i
v
e m
od
el
i
n
g ar
e
v
er
y
l
i
m
i
t
ed.
Ke
y
w
o
rd
s
:
r
et
w
e
et
pr
edi
c
t
i
on,
m
ac
hi
n
e l
e
ar
ni
ng a
l
gor
i
t
hm
s
,
per
f
or
m
an
c
e
C
o
p
y
r
i
g
h
t
©
20
16 U
n
i
ver
si
t
a
s A
h
mad
D
ah
l
an
.
A
l
l
r
i
g
h
t
s r
eser
ved
.
1
.
I
n
tr
o
d
u
c
ti
o
n
S
oc
i
a
l
m
edi
a
l
i
k
e T
w
i
t
t
er
ha
s
pr
ov
i
de
d a
p
l
at
f
or
m
f
or
s
pr
eadi
ng
i
nf
or
m
at
i
on am
ong
us
er
s
[
1,
2]
.
I
n
t
hi
s
w
or
k
w
e
f
oc
us
on
t
he
r
et
w
e
et
pr
e
di
c
t
i
on
pr
obl
em
.
G
i
v
en
a
t
w
e
et
,
w
e
w
o
u
ld
lik
e
t
o
pr
edi
c
t
w
h
et
h
er
i
t
w
i
l
l
b
e
r
et
w
eet
ed.
A
pp
l
i
c
at
i
ons
of
t
hi
s
t
as
k
ar
e,
f
or
ex
a
m
pl
e,
t
o
hel
p
dec
i
s
i
o
n
m
a
k
er
s
pr
opagat
e t
h
ei
r
i
s
s
ues
an
d f
ac
i
l
i
t
at
e
c
om
pani
e
s
t
o pr
om
ot
e t
he
i
r
pr
od
uc
t
s
.
A
w
i
d
e
r
ange
of
l
ear
n
i
ng
m
et
hods
and
f
eat
ur
es
hav
e
b
een
pr
o
po
s
ed
f
or
r
et
w
ee
t
pr
edi
c
t
i
on
;
s
ee,
e.
g.
,
[3
-
7]
.
I
n a
ddi
t
i
o
n,
di
f
f
er
ent
m
odel
i
ng s
et
ups
h
av
e b
een
pr
opos
ed
;
e.
g.
,
P
e
t
r
o
vi
c
,
e
t
a
l.
,
[
3]
pr
opos
e a t
i
m
e
-
s
ens
i
t
i
v
e m
odel
t
hat
b
ui
l
ds
s
ep
ar
at
e
m
odel
s
depe
ndi
ng
on
t
he
t
w
eet
'
s
c
r
eat
i
on
t
i
m
e
and
s
ho
w
t
hat
i
t
s
ubs
t
ant
i
al
l
y
i
m
pr
ov
es
per
f
or
m
anc
e
ov
er
t
he
p
as
s
i
v
e
-
aggr
es
s
i
v
e
l
e
ar
ni
ng
al
g
or
i
t
hm
.
T
hi
s
l
ar
ge
v
ar
i
et
y
of
f
eat
ur
es
,
l
e
ar
ni
ng m
et
hods
,
and
w
a
y
s
of
m
odel
i
ng t
he t
as
k
c
al
l
s
f
or
a s
y
s
t
em
at
i
c
c
o
m
par
i
s
on o
n a s
i
ng
l
e d
at
as
et
.
T
her
ef
o
r
e,
w
e pr
op
os
e
a s
y
s
t
em
at
i
c
c
o
m
par
i
s
on
of
t
he s
el
ec
t
e
d l
ear
ni
n
g m
et
hods
and f
eat
ur
es
,
bo
t
h
w
i
t
h
i
n an
d w
i
t
h
out
t
im
e
-
s
ens
i
t
i
v
e
f
r
am
ew
or
k
,
al
l
o
n a
s
i
n
gl
e d
at
as
et
.
T
o t
h
e a
ut
h
or
s
’
k
no
w
l
edg
e
t
hi
s
i
s
t
h
e f
i
r
s
t
at
t
em
pt
t
o c
ol
l
ec
t
b
es
t
p
er
f
or
m
i
ng f
eat
ur
es
and c
o
nt
r
a
s
t
l
i
n
ear
a
nd
non
-
l
i
n
ear
l
ear
ni
n
g m
et
hods
.
W
e
c
ons
i
der
t
he
f
ol
l
o
w
i
ng
r
es
ear
c
h
ques
t
i
ons
:
(
i
)
W
hi
c
h
of
t
he
pr
opos
ed
l
e
ar
ni
n
g
m
et
hods
i
s
t
he
m
os
t
ef
f
ec
t
i
v
e f
or
t
he r
et
w
eet
pr
ed
i
c
t
i
o
n t
as
k
?
(
i
i
)
W
hi
c
h o
f
t
he pr
opos
ed f
eat
ur
es
ar
e m
os
t
di
s
c
r
i
m
i
nat
i
v
e f
eat
ur
es
f
or
t
he l
e
ar
ni
ng m
et
ho
ds
c
ons
i
der
e
d
?
A
nd (
i
i
i
)
T
o w
hi
c
h d
egr
e
e
does
t
i
m
e
-
s
ens
i
t
i
v
e
m
odel
i
ng
h
el
p i
m
pr
ov
e t
he p
er
f
or
m
anc
e
of
l
ear
ni
ng
m
et
hods
on
o
ur
dat
as
et
?
P
r
i
or
w
or
k
has
doc
um
ent
e
d s
ev
er
al
t
ec
hn
i
q
ues
t
o s
ol
v
e t
he r
et
w
e
et
pr
edi
c
t
i
on
pr
obl
em
.
N
av
eed
,
e
t
a
l.
,
[
5
]
s
t
at
e
t
h
at
t
he
pr
ob
l
em
of
f
i
ndi
ng
"
i
nt
er
es
t
i
ngn
es
s
"
on
T
w
i
t
t
er
i
s
t
he
s
am
e
as
pr
edi
c
t
i
ng
w
h
et
her
a
t
w
e
et
w
i
l
l
b
e
r
et
w
eet
ed.
T
hey
em
pl
o
y
l
og
i
s
t
i
c
r
e
gr
es
s
i
on
t
o
d
o
t
he
pr
edi
c
t
i
on
a
nd
f
i
nd
t
hat
c
on
t
ent
f
eat
ur
es
s
uc
h
as
i
ni
t
i
al
negat
i
v
e
s
ent
i
m
ent
s
m
a
k
e
a
t
w
ee
t
m
or
e
l
i
k
el
y
t
o b
e r
et
w
eet
e
d.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
C
o
mpar
i
s
on
of
R
e
t
w
ee
t
P
r
ed
i
c
t
i
o
n A
ppr
oac
hes
:
T
h
e
S
u
per
i
or
i
t
y
…
(
H
endr
a
B
u
ny
ami
n
)
1053
P
e
t
r
o
vi
c
,
e
t
a
l.
,
[
3]
c
as
t
t
he
r
et
w
eet
pr
ed
i
c
t
i
o
n pr
o
bl
em
as
a bi
nar
y
c
l
as
s
i
f
i
c
at
i
on pr
obl
em
w
it
h
a
t
im
e
-
s
ens
i
t
i
v
e
m
odel
i
ng
a
ppr
o
ac
h.
T
he
y
ar
gu
e
t
hat
t
i
m
e
-
s
ens
i
t
i
v
e
m
odel
i
ng
s
ubs
t
ant
i
al
l
y
hel
ps
i
m
pr
ov
e
per
f
or
m
anc
e
ov
er
p
as
s
i
v
e
-
aggr
es
s
i
v
e
al
gor
i
t
hm
on
t
h
e
i
r
d
at
as
et
.
T
he
y
a
l
s
o c
l
ai
m
t
hat
s
oc
i
a
l
f
eat
ur
es
t
hat
ar
e r
el
at
ed t
o a us
er
i
m
pr
ov
e t
he ac
c
ur
ac
y
of
t
hei
r
m
odel
m
or
e t
han
t
w
eet
f
eat
ur
es
.
O
ur
w
or
k
ut
i
l
i
z
es
t
he t
i
m
e
-
s
ens
i
t
i
v
e (
T
S
)
m
odel
i
ng;
ho
w
ev
er
,
t
he
T
S
m
odel
i
ng
i
n
out
d
at
as
et
has
l
i
m
i
t
ed c
on
t
r
i
but
i
on
f
or
o
ur
pr
ed
i
c
t
i
on t
a
s
k
.
P
r
ed
i
c
t
i
n
g
r
et
w
eet
s
and
r
epl
i
es
f
or
a
g
i
v
en
t
w
ee
t
al
s
o
don
e
b
y
A
r
t
z
i
,
et
a
l
.
,
[6
]
.
S
pec
i
f
i
c
a
l
l
y
,
f
or
a gi
v
en t
w
eet
,
t
h
e
y
di
s
c
ov
er
t
hat
r
em
ov
i
ng s
oc
i
al
f
eat
ur
es
s
uc
h
as
num
ber
o
f
f
ol
l
o
w
er
s
,
n
um
ber
of
f
ol
l
o
w
ees
,
and r
a
t
i
o
bet
w
een
t
h
e t
w
o c
aus
es
a b
i
g dr
op i
n t
he
i
r
m
odel
'
s
pr
edi
c
t
i
on
ac
c
ur
ac
y
o
v
er
m
ode
l
s
t
hat
i
nc
l
ude
t
hem
.
C
ar
uan
a
,
e
t
a
l.
,
[
8]
m
ent
i
on
r
andom
f
or
es
t
-
bas
ed l
e
ar
ni
ng m
et
hod w
i
t
hout
c
al
i
br
at
i
o
n gi
v
e t
he b
e
s
t
av
er
ag
e per
f
or
m
anc
e ac
r
os
s
al
l
m
et
r
i
c
s
and t
es
t
pr
ob
l
em
s
;
m
or
eo
v
e
r
,
F
er
na
nde
z
-
D
e
l
ga
do
,
e
t
a
l.
,
[
9]
c
ons
t
r
uc
t
t
hor
oug
h ex
per
i
m
ent
s
and c
onc
l
ude r
an
dom
f
or
es
t
-
bas
ed l
e
ar
ni
ng m
et
hod ac
hi
e
v
e t
h
e m
ax
i
m
u
m
ac
c
ur
a
c
y
.
T
her
ef
or
e
,
w
e
op
t
t
o
ut
i
l
i
z
e
r
and
om
f
o
r
es
t
-
bas
ed
l
ear
n
i
ng
m
et
hod
r
at
her
t
ha
n em
pl
o
y
M
A
R
T
i
n
our
w
or
k
.
Mor
eo
v
er
,
i
t
t
ur
ns
out
t
h
at
r
andom
f
or
es
t
-
bas
ed l
ear
ni
n
g
m
et
hod,
w
h
i
c
h has
no
t
b
een em
pl
o
y
e
d
i
n pr
e
v
i
ous
s
t
u
di
es
,
gi
v
es
s
uper
i
or
r
es
ul
t
s
i
n
our
c
om
par
i
s
on s
t
u
d
y
.
H
ong
,
e
t
a
l.
,
[
4]
pr
e
di
c
t
w
h
et
her
or
not
a
t
w
e
et
w
i
l
l
be
r
et
w
eet
ed
an
d
ho
w
m
an
y
t
i
m
e
s
a
t
w
eet
w
i
l
l
be
r
et
w
e
et
e
d.
T
he
y
em
pl
o
y
l
o
gi
s
t
i
c
r
egr
es
s
i
on
i
n
t
he
i
r
w
or
k
.
H
o
w
m
any
t
i
m
es
a
ne
w
t
w
eet
w
i
l
l
be r
e
t
w
ee
t
ed
ba
s
ed on
a c
er
t
a
i
n
t
hr
es
ho
l
d
i
s
s
t
ud
i
ed
b
y
J
en
der
s
,
et
al
.
,
[
10]
.
T
he
y
ut
i
l
i
z
e
m
odel
s
s
uc
h
as
N
a
i
v
e
-
B
a
y
es
a
nd
gen
er
al
i
z
ed
l
i
n
ear
m
odel
.
O
t
h
er
w
or
k
b
y
G
a
o
,
et
al
.
,
[
11]
a
l
s
o pr
ed
i
c
t
ho
w
m
any
t
i
m
es
a t
w
e
et
w
i
l
l
r
et
w
e
et
ed b
y
a
ppl
y
i
ng a
n ex
t
en
ded r
ei
nf
or
c
ed
P
oi
s
s
on
pr
oc
es
s
m
odel
w
i
t
h t
i
m
e m
appi
ng pr
oc
es
s
.
Z
am
an
,
et
al
.
,
[
1
2]
m
eas
ur
es
t
he p
opu
l
ar
i
t
y
of
a t
w
eet
t
hr
oug
h t
h
e t
i
m
e
-
s
er
i
es
pat
h
of
i
t
s
r
et
w
eet
s
.
A
B
a
y
es
i
an
pr
ob
abi
l
i
s
t
i
c
m
odel
i
s
de
v
el
ope
d f
or
t
he
ev
ol
u
t
i
o
n of
t
h
e
r
et
w
eet
s
an
d
popu
l
ar
i
t
y
of
a t
w
eet
i
s
pr
edi
c
t
e
d bas
ed o
n t
he r
et
w
eet
t
i
m
es
and l
oc
al
n
et
w
or
k
or
"
gr
aph
"
s
t
r
uc
t
ur
e
of
r
et
w
eet
er
s
.
Mac
s
k
as
s
y
,
e
t
a
l.
,
[
13
]
t
a
g
t
w
eet
s
w
i
t
h
W
i
k
i
pedi
a
c
at
egor
i
es
an
d
gener
at
e pr
of
i
l
es
of
"
t
o
pi
c
s
of
i
nt
er
es
t
"
b
as
ed o
n pa
s
t
c
ont
ent
p
os
t
ed
and c
o
n
s
t
r
uc
t
r
et
w
eet
beha
v
i
or
m
odel
s
f
or
us
er
s
.
T
he
y
ar
g
ue t
h
at
pe
opl
e'
s
r
et
w
eet
i
ng b
eha
v
i
or
i
s
be
t
t
er
ex
pl
a
i
n
ed
t
hr
oug
h m
ul
t
i
p
l
e d
i
f
f
er
ent
m
ode
l
s
r
at
her
t
han
on
e m
odel
.
Mor
c
hi
d
,
et
a
l.
,
[
14]
s
t
ud
y
t
he beh
av
i
or
of
t
w
e
et
s
t
hat
hav
e bee
n
m
a
ssi
ve
l
y
r
et
w
e
et
ed i
n
a s
hor
t
t
i
m
e.
S
p
ec
i
f
i
c
al
l
y
,
t
he
y
em
pl
o
y
P
r
i
nc
i
pa
l
C
om
ponent
A
nal
y
s
i
s
t
o s
el
ec
t
f
eat
ur
es
.
C
om
par
ed t
o o
ur
w
or
k
,
t
he
y
ex
t
r
ac
t
l
es
s
num
ber
of
f
e
at
ur
es
an
d n
um
ber
of
l
ear
ni
ng m
et
hods
.
Xu
,
et
al
.
,
[
7]
an
al
y
z
e
us
er
r
et
w
eet
b
eha
v
i
or
at
i
nd
i
v
i
d
ual
l
ev
el
an
d
ar
gue
t
h
at
t
h
e
m
os
t
i
m
por
t
ant
f
eat
ur
es
f
or
gener
al
pe
op
l
e ar
e s
oc
i
al
.
O
u
r
w
or
k
i
s
s
i
m
i
l
ar
t
o t
he
i
r
s
;
ho
w
e
v
er
,
ou
r
f
oc
us
i
s
s
pec
i
f
i
c
al
l
y
t
o
und
er
s
t
and
t
w
ee
t
s
f
r
o
m
pol
i
t
i
c
i
ans
an
d
w
e
em
pl
o
y
m
or
e
al
g
or
i
t
hm
s
t
ha
t
hav
e
not
bee
n
t
es
t
e
d,
s
pec
i
f
i
c
al
l
y
r
an
dom
f
or
es
t
-
bas
ed
l
ear
ni
n
g
m
et
hod,
an
d
m
o
r
e
m
odel
i
ng
i
n
our
ex
per
i
m
ent
s
.
Mor
eo
v
e
r
,
t
he r
and
om
f
or
es
t
-
bas
ed l
ear
ni
n
g m
et
hod t
hat
w
e
em
pl
o
y
gi
v
es
bet
t
er
pr
ed
i
c
t
i
o
n
ac
c
ur
ac
y
t
han ot
her
al
gor
i
t
hm
s
t
h
ey
em
p
l
oy
.
2.
M
o
d
e
l
i
n
g
a
n
d
F
e
a
tu
r
e
s
I
n t
hi
s
s
ec
t
i
on
w
e
des
c
r
i
be
our
m
odel
i
n
g appr
oac
h f
or
addr
es
s
i
n
g t
he r
e
t
w
ee
t
pr
e
di
c
t
i
on
pr
obl
em
and c
as
t
t
he
pr
ob
l
em
as
a bi
nar
y
c
l
as
s
i
f
i
c
at
i
o
n pr
ob
l
em
.
2
.1
.
M
o
d
e
l
i
n
g
fo
r
R
e
tw
e
e
t P
r
e
d
i
c
ti
o
n
T
abl
e 1 d
es
c
r
i
bes
4
gr
ou
p
s
of
l
ear
ner
s
t
hat
w
e em
pl
o
y
i
n
our
m
odel
i
ng a
ppr
o
ac
h.
T
abl
e 1.
Lear
n
i
n
g a
ppr
oac
h
es
.
G
r
oup of
l
ear
ner
s
M
odel
G
l
obal
l
i
near
G
l
obal
pas
s
i
v
e
-
aggr
es
s
i
v
e
(
G
-
PA)
G
l
obal
l
i
near
s
uppor
t
v
e
c
t
or
(
G
-
L
SV
)
G
l
obal
l
ogi
s
t
i
c
r
egr
es
s
i
on
(
G
-
L
R)
G
l
obal
non
-
l
i
near
G
l
obal
dec
i
s
i
on
t
r
ee (
G
-
DT
)
G
l
obal
r
ando
m
f
or
es
t
(
G
-
RF
)
T
S
p
a
s
si
ve
-
aggr
es
s
i
v
e (
T
S
-
PA
)
T
im
e
-
s
en
s
i
t
i
v
e
(
T
S
)
l
i
near
T
S
l
i
near
s
uppor
t
v
ec
t
or
(
T
S
-
L
SV)
T
S
l
ogi
s
t
i
c
r
egr
es
s
i
on
(
T
S
-
L
R)
T
im
e
-
s
en
s
i
t
i
v
e
(
T
S
)
non
-
l
i
near
T
S
de
c
i
s
i
on
t
r
ee (
T
S
-
DT
)
T
S
r
ando
m
f
or
es
t
(
T
S
-
RF
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
52
–
1
058
1054
T
he c
l
as
s
i
f
i
c
at
i
on r
u
l
e f
or
a
gl
o
bal
l
i
ne
ar
l
e
ar
ner
i
s
:
=
s
ig
n
(
〈
,
〉
)
,
(
1)
W
he
r
e
i
s
t
he
g
l
o
bal
w
ei
ght
v
ec
t
or
,
i
s
t
he
f
eat
ur
e
v
ec
t
or
r
epr
es
ent
a
t
i
o
n
of
a
t
w
ee
t
,
and,
is
t
he pr
e
di
c
t
i
on
.
T
he c
l
as
s
i
f
i
c
at
i
o
n r
ul
e f
or
a g
l
ob
al
no
n
-
l
i
near
l
ear
n
er
i
s
:
=
a
rg
ma
x
∈
{
0
,
1
}
(
(
|
)
)
,
(
2)
W
h
er
e
i
s
a g
l
o
bal
t
r
ee
m
ode
l
a
nd
i
s
t
h
e pr
e
di
c
t
e
d
c
l
as
s
det
er
m
i
ned
as
t
he
m
ax
i
m
u
m
a
pos
t
er
i
or
(
MA
P
)
of
t
he c
l
as
s
di
s
t
r
i
bu
t
i
on t
he
f
a
lls
in
t
h
e
le
a
f
.
T
im
e
-
s
ens
i
t
i
v
e (
T
S
)
m
odel
i
ng [
2]
as
s
um
es
t
hat
t
her
e ar
e s
om
e s
pec
i
f
i
c
r
ul
es
w
i
t
h
i
n
ev
er
y
ho
ur
of
a da
y
f
or
t
w
eet
s
be
i
n
g r
et
w
e
et
ed
.
E
ac
h
hour
i
n a d
a
y
c
or
r
es
p
onds
t
o a
l
oc
a
l
m
odel
,
t
her
ef
or
e,
T
S
m
odel
i
ng c
o
ns
i
s
t
s
of
one
g
l
ob
al
m
odel
(
ei
t
h
er
eq
uat
i
on
(
1)
or
equ
at
i
on (
2)
)
and
24
l
oc
al
m
odel
s
.
T
he T
S
l
i
ne
ar
l
ear
ner
i
s
t
he
n
:
=
s
ig
n
〈
,
〉
+
λ
〈
,
〉
,
(
3)
W
h
er
e,
i
s
t
he
l
oc
al
w
e
i
gh
t
v
ec
t
or
a
nd
λ
i
s
t
h
e
w
ei
gh
t
t
h
at
c
or
r
es
po
nds
t
o t
he
num
ber
of
t
w
eet
s
t
h
at
t
h
e l
oc
al
m
odel
has
s
een d
ur
i
n
g t
r
a
i
ni
ng,
di
v
i
ded
b
y
t
he t
ot
a
l
num
ber
o
f
t
w
eet
s
i
n t
h
e
t
r
ai
n
i
ng s
et
.
F
i
na
l
l
y
,
t
he T
S
non
-
l
i
n
ear
l
ear
ner
i
s
d
ef
i
n
e
d as
:
=
a
rg
ma
x
∈
{
0
,
1
}
(
(
|
)
+
λ
(
|
)
)
,
(
4)
W
h
er
e
i
s
a l
oc
a
l
t
r
e
e m
odel
.
2
.
2
.
L
e
a
r
n
i
n
g
M
e
th
o
d
s
B
as
ed
on pr
e
v
i
ous
s
t
u
di
es
(
[
3
-
5
],
[7
,
10]
)
,
w
e pr
op
o
s
e 5 l
e
ar
ni
ng m
et
hods
:
pa
s
s
i
v
e
-
aggr
es
i
v
e
(
P
A
)
,
l
i
near
s
up
por
t
v
ec
t
or
m
ac
hi
ne (
LS
V
)
,
l
ogi
s
t
i
c
r
egr
es
s
i
on (
LR
)
,
dec
i
s
i
o
n t
r
ees
(
D
T
)
and r
andom
f
or
es
t
-
bas
ed
l
ear
n
i
n
g (
R
F
)
.
C
om
bi
ne
d
w
i
t
h
t
he
c
hoi
c
e
f
or
gl
oba
l
v
s
.
t
i
m
e
-
s
ens
i
t
i
v
e m
odel
i
n
g,
t
h
i
s
y
i
e
l
ds
a t
ot
al
of
10 a
ppr
oac
h
es
;
s
ee T
abl
e
1.
2
.
3
.
D
esc
r
i
p
ti
o
n
o
f F
e
a
tu
r
e
s
Li
k
e
t
he
l
e
ar
ni
ng
a
ppr
oac
h
es
,
t
he
f
eat
ur
es
t
h
at
w
e
c
o
ns
i
der
ar
e
bas
ed
on
a
n
um
ber
of
pr
ev
i
ous
s
t
u
di
es
;
s
ee T
abl
e
2.
T
abl
e
2
.
F
eat
ur
es
and
t
he
i
r
or
i
g
i
ns
.
U
ser
f
eat
u
r
es
N
um
ber
o
f
f
ol
l
ow
er
s
[1
5
],
[3
],
[4
]
, [6
]
,
[7
]
, [1
0
],
[1
2
]
N
um
ber
o
f
f
r
i
ends
[1
5
]
,
[3
],
[4
]
, [6
]
,
[7
]
N
um
ber
o
f
s
t
a
t
us
e
s
[1
5
]
,
[3
],
[7
]
N
um
ber
o
f
f
av
or
i
t
es
[1
5
]
,
[3
]
N
um
ber
o
f
u
s
er
l
i
s
t
ed
[3
],
[
7
]
I
s
a u
s
er
v
er
i
f
i
ed?
[3
],
[
7
]
P
er
c
ent
age
o
f
r
epl
i
e
s
[
6]
N
um
ber
o
f
f
ol
l
ow
er
s
/
N
u
m
ber
of
f
r
i
ends
[
6]
A
v
er
age num
ber
of
t
w
eet
s
per
day
[7
],
[
1
3
]
A
c
c
ount
age
[1
5
]
,
[7
]
T
w
eet
f
eat
u
r
es
N
um
ber
o
f
ha
s
ht
ag
s
[1
5
]
,
[3
],
[6
]
, [7
]
,
[1
0
]
N
um
ber
o
f
m
ent
i
on
s
[1
5
]
,
[3
],
[6
]
, [7
]
,
[1
0
]
Nu
m
b
e
r
o
f
URL
s
[4
],
[
1
0
]
Lengt
h of
a
t
w
eet
[3
],
[
6
],
[1
0
]
N
ov
el
t
y
s
c
or
e
[3
],
[
3
]
I
s
a t
w
eet
a
r
epl
y
?
[3
],
[
7
]
I
s
a t
w
eet
a
di
r
ec
t
m
e
s
s
age?
[
5]
D
oes
a
t
w
eet
c
ont
ai
n
a ha
s
ht
ag?
[
5]
D
oes
a
t
w
eet
c
ont
ai
n
a U
R
L?
[
5]
D
oes
a
t
w
eet
c
ont
ai
n
‘
?’
or
‘
!
’
?
[
5]
W
e
di
v
i
de t
he f
eat
ur
es
i
nt
o
2 c
at
eg
or
i
es
as
f
ol
l
o
w
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
C
o
mpar
i
s
on
of
R
e
t
w
ee
t
P
r
ed
i
c
t
i
o
n A
ppr
oac
hes
:
T
h
e
S
u
per
i
or
i
t
y
…
(
H
endr
a
B
u
ny
ami
n
)
1055
2.
3.
1
.
U
s
e
r F
e
a
tu
r
e
s
A
us
er
c
an
pos
s
es
s
at
t
r
i
bu
t
es
t
hat
m
a
k
e
her
t
w
eet
m
or
e
l
i
k
el
y
t
o
be
r
et
w
e
et
e
d.
W
e
t
r
y
t
o
c
apt
ur
e
t
hos
e
at
t
r
i
b
ut
es
f
r
o
m
i
nf
or
m
at
i
on
abou
t
t
he
us
er
as
f
ol
l
o
w
s
:
ho
w
m
an
y
peopl
e
f
ol
l
o
w
t
he us
er
(
n
um
ber
of
f
ol
l
ow
er
s
)
,
ho
w
m
an
y
pe
op
l
e t
h
e us
er
f
ol
l
o
w
s
(
num
ber
of
f
r
i
ends
)
,
ho
w
m
an
y
s
t
a
t
us
es
t
he
us
er
h
as
(
num
ber
of
s
t
at
us
es
)
,
ho
w
m
an
y
f
a
v
or
i
t
e t
w
eet
s
t
he
us
er
has
(
num
ber
of
f
av
or
i
t
es
)
,
ho
w
m
an
y
t
i
m
es
t
he
us
er
i
s
l
i
s
t
ed
(
num
ber
of
us
er
l
i
s
t
ed)
,
w
het
h
er
or
not
t
he
us
er
i
s
v
er
i
f
i
e
d (
i
s
a u
s
er
v
er
i
f
i
e
d?)
.
F
r
om
al
l
t
w
e
e
t
s
aut
hor
ed
b
y
a
us
er
,
w
e c
om
put
e t
he
r
at
i
o
of
t
w
e
et
s
t
h
at
h
av
e r
e
pl
i
es
t
o a
l
l
her
t
w
eet
s
(
per
c
ent
a
ge of
r
ep
l
i
es
)
,
pr
opor
t
i
o
n of
num
ber
of
f
ol
l
o
w
er
s
t
o
n
um
ber
of
f
r
i
ends
(
num
ber
of
f
ol
l
o
w
er
s
/
num
ber
of
f
r
i
ends
)
,
ho
w
m
an
y
t
w
eet
s
t
h
e
us
er
publ
i
s
he
d on a
v
er
a
g
e per
da
y
(
a
v
er
age n
um
ber
of
t
w
eet
s
per
d
a
y
)
,
a
n
d ho
w
o
l
d her
ac
c
ount
a
ge
of
a us
er
i
s
(
i
n
da
y
s
)
w
hen s
he p
ub
l
i
s
he
d
t
he t
w
eet
(
ac
c
ou
nt
age)
.
2.
3.
2
.
T
w
eet
F
eat
u
r
es
W
e
al
s
o
ex
t
r
ac
t
f
eat
ur
es
about
a
nd
f
r
om
t
w
ee
t
s
t
hem
s
el
v
es
.
W
e
c
ons
i
der
onl
y
f
eat
ur
es
t
hat
h
a
v
e
be
en
s
h
o
w
n
i
m
p
or
t
ant
i
n
t
he
ex
i
s
t
i
ng
w
or
k
.
T
he
f
eat
ur
es
f
r
o
m
t
w
eet
s
ar
e
as
f
ol
l
o
w
s
:
num
ber
of
has
ht
ags
,
m
ent
i
ons
,
U
R
Ls
,
l
engt
h of
t
h
e t
w
eet
,
t
he
no
v
e
l
t
y
s
c
or
e,
w
he
t
her
or
not
t
h
e
t
w
eet
i
s
a
r
ep
l
y
,
a
di
r
ec
t
m
e
s
s
age,
w
he
t
her
or
no
t
t
he t
w
eet
c
o
nt
a
i
ns
a h
a
s
ht
ag,
a
U
R
L,
a
n
ex
c
l
am
at
i
on,
or
q
ues
t
i
on m
ar
k
s
.
N
ov
el
t
y
s
c
or
e
i
s
c
om
put
e
d as
t
he
c
os
i
n
e d
i
s
t
a
nc
e be
t
w
e
en t
he
T
F
-
I
D
F
v
ec
t
or
r
epr
es
ent
at
i
ons
of
t
he t
w
eet
a
nd
i
t
s
near
es
t
n
ei
ghb
or
t
w
ee
t
pu
bl
i
s
h
ed a
da
y
bef
or
e.
3.
E
xp
er
i
m
en
t
al
S
et
u
p
I
n or
der
t
o und
er
s
t
and
w
h
a
t
k
i
nd of
t
w
eet
s
w
ou
l
d be r
e
t
w
eet
ed,
w
e c
r
eat
e
d a dat
a
s
et
as
f
o
llo
w
s
.
W
e c
ol
l
ec
t
ed l
i
s
t
s
of
D
ut
c
h pol
i
t
i
c
i
ans
a
nd p
o
l
i
t
i
c
al
j
our
nal
i
s
t
s
f
r
o
m
l
i
s
t
s
c
ur
at
ed b
y
De
I
s
s
uemak
er
s
,
a D
u
t
c
h c
o
m
m
uni
c
at
i
ons
c
ons
u
l
t
a
nc
y
c
o
m
pan
y
,
al
o
ng
w
i
t
h t
he
i
r
f
ol
l
o
w
er
s
an
d
f
o
llo
w
e
es
.
T
he t
ot
al
num
ber
of
pol
i
t
i
c
i
ans
a
nd j
our
na
l
i
s
t
s
i
s
3
04
and
t
ot
a
l
n
um
ber
of
f
ol
l
o
w
er
s
and f
ol
l
o
w
e
es
i
s
ar
oun
d 1.
4 m
i
l
l
i
on
.
W
e c
ol
l
ec
t
ed b
ot
h t
w
eet
s
a
nd us
er
pr
of
i
l
es
f
r
om
S
ept
em
ber
9
t
o
D
ec
em
ber
2,
2
01
4
b
y
ut
i
l
i
z
i
ng
t
he
T
w
i
t
t
er
A
P
I
.
W
e
gat
her
ed
ar
ou
nd
3
m
i
l
l
i
on
t
w
e
et
s
.
O
ur
t
r
ai
n
i
ng
s
et
c
om
pr
i
s
es
t
w
ee
t
s
f
r
o
m
S
ept
em
ber
9
t
o
D
e
c
e
m
ber
1
(
ar
ound
2.
7
m
i
l
l
i
o
n
t
w
eet
s
)
an
d
w
e us
e t
w
eet
s
f
r
om
t
he l
as
t
da
y
,
D
ec
em
ber
2nd
as
our
t
es
t
s
et
(
ar
o
und
0.
3 m
i
l
l
i
o
n
t
w
eet
s
)
.
W
e
l
abel
a t
w
eet
as
r
et
w
e
et
ed
or
not
r
et
w
eet
ed
b
y
c
hec
k
i
ng
w
het
her
t
he
t
w
ee
t
has
a
or
i
gi
na
l
s
t
at
us
i
d.
I
f
i
t
h
as
o
ne,
w
e ac
q
ui
r
e
t
h
e or
i
gi
na
l
t
w
eet
w
i
t
h t
he s
t
at
us
i
d
and
gi
v
e
i
t
l
a
be
l
1
.
I
f
i
t
has
n
o
or
i
gi
na
l
s
t
at
us
i
d,
w
e
l
a
be
l
t
h
e
t
w
e
et
0.
W
e
s
et
a
t
hr
es
ho
l
d
v
al
u
e
t
o
2
da
y
s
t
o
gi
v
e
a
t
w
eet
a c
ha
nc
e t
o
be r
e
t
w
e
et
ed.
T
he pr
opor
t
i
o
n of
r
et
w
eet
e
d t
w
eet
s
o
v
er
a
l
l
i
s
ar
ound
33%
.
T
abl
e
3
.
T
he b
es
t
s
et
t
i
ng
s
af
t
er
5
-
f
ol
d c
r
os
s
-
v
al
i
dat
i
o
n
No
M
ode
l
S
e
tti
n
g
1
G
-
PA,
T
S
-
PA
C
=
0.
01,
l
os
s
=
s
quar
ed
-
hi
nge
2
G
-
L
SV,
T
S
-
LS
V
C
=
10,
dual
=
f
al
s
e
3
G
-
L
R,
T
S
-
LR
C
=
10,
penal
t
y
=
l
2
4
G
-
D
T
,
TS
-
DT
c
r
i
t
er
i
on =
ent
r
opy
,
s
pl
i
t
t
er
=
be
s
t
5
G
-
R
F
,
TS
-
RF
c
r
i
t
er
i
on =
gi
ni
,
n_es
t
i
m
at
or
s
=
30
T
abl
e
4
.
F
1
(
%
)
s
c
or
es
af
t
er
5
-
f
ol
d c
r
os
s
-
v
al
i
da
t
i
o
n o
n t
he t
r
a
i
ni
ng s
e
t
M
ode
l
A
v
er
ag
e
±
s
td
M
ode
l
A
v
er
ag
e
±
s
td
G
l
obal
m
odel
s
T
S
m
odel
s
G
-
PA
5
1
.
8
1
±
0
.
0
2
2
9
TS
-
PA
5
2
.
5
4
±
0
.
0
1
6
0
G
-
LS
V
5
6
.
4
4
±
0
.
0
0
1
0
TS
-
L
SV
5
6
.
4
9
±
0
.
0
0
1
1
G
-
LR
5
6
.
6
0
±
0
.
0
0
1
1
TS
-
LR
5
6
.
6
6
±
0
.
0
0
1
1
G
-
DT
6
9
.
1
8
±
0
.
0
0
0
9
TS
-
DT
6
9
.
7
5
±
0
.
0
0
1
2
G
-
RF
7
4
.
3
9
±
0
.
0
0
0
6
TS
-
RF
7
5
.
4
2
±
0
.
0
0
0
8
B
ef
or
e w
e r
un t
he pr
e
di
c
t
i
o
n on t
he t
es
t
s
et
,
w
e r
un 5
-
f
ol
d c
r
os
s
-
v
al
i
da
t
i
o
n
w
i
t
h s
el
ec
t
ed
s
et
t
i
n
gs
[
16,
17]
f
or
eac
h c
l
as
s
i
f
i
er
on our
t
r
ai
n
i
ng s
et
.
T
he pur
pos
e of
c
r
os
s
-
v
al
i
d
at
i
o
n i
s
t
o t
un
e
al
l
t
he m
odel
s
and f
i
nd t
h
e bes
t
s
et
t
i
ng f
r
o
m
eac
h m
odel
.
T
abl
e 3 des
c
r
i
bes
t
he bes
t
s
et
t
i
n
g
r
es
ul
t
i
ng f
r
om
t
he c
r
os
s
-
v
al
i
dat
i
on
on t
h
e t
r
ai
ni
ng s
et
.
T
he c
l
as
s
i
f
i
c
at
i
on
per
f
or
m
anc
e of
our
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
52
–
1
058
1056
m
odel
s
af
t
er
c
r
os
s
-
v
al
i
d
at
i
on
i
s
s
ho
w
n
i
n
T
abl
e
4.
W
e
s
ee
t
hat
t
he
n
on
-
l
i
ne
ar
m
ode
l
s
(
D
T
,
R
F
)
out
p
er
f
or
m
t
he l
i
ne
ar
m
odel
s
(
P
A
,
LS
V
,
L
R
)
,
and
t
h
at
t
i
m
e
-
s
ens
i
t
i
v
e
m
odel
i
ng
onl
y
m
ar
gi
nal
l
y
out
p
er
f
or
m
s
gl
obal
m
odel
i
n
g on
t
he
t
r
ai
n
i
n
g s
et
.
F
or
s
i
gni
f
i
c
anc
e t
es
t
i
n
g,
w
e us
e a on
e
-
t
ai
l
ed
p
ai
r
e
d t
-
t
es
t
f
or
c
o
m
par
i
s
ons
bet
w
een
gl
o
bal
l
e
ar
ner
s
an
d
be
t
w
ee
n
t
i
m
e
-
s
ens
i
t
i
v
e
l
ear
n
er
s
;
s
i
gn
i
f
i
c
ant
d
i
f
f
er
enc
es
ar
e
m
ar
k
ed
us
i
ng
↑
f
or
s
i
gni
f
i
c
ant
di
f
f
er
enc
es
at
=
0
.
0
1
.
W
e us
e Mc
N
e
m
ar
'
s
t
e
s
t
t
o
m
eas
ur
e t
he s
i
gn
i
f
i
c
anc
e
di
f
f
er
enc
es
i
n pr
e
di
c
t
i
on
ac
c
ur
ac
y
of
t
he
gl
o
ba
l
l
ear
n
er
s
and t
he T
S
l
ear
ner
s
.
4
.
R
e
s
u
l
ts
4.
1.
P
r
ed
i
ct
i
o
n
A
cc
u
r
acy
W
e
r
un
al
l
m
odel
s
f
r
om
T
a
bl
e
1.
T
abl
e
5
s
h
o
w
s
t
he
F
1
s
c
or
e of
t
he m
odel
s
w
i
t
h
us
er
f
eat
ur
es
on
l
y
,
t
w
e
et
f
eat
ur
e
s
onl
y
,
and
w
i
t
h b
ot
h
of
t
he
m
on t
he t
es
t
s
e
t
.
W
e s
ee t
hat
t
he
gl
o
ba
l
and
t
i
m
e
-
s
ens
i
t
i
v
e
r
an
dom
f
or
es
t
m
odel
(
G
-
R
F
)
and
(
T
S
-
R
F
)
ac
hi
ev
e
t
h
e
hi
ghes
t
per
f
or
m
anc
e.
W
e
al
s
o s
ee t
h
at
us
er
f
eat
u
r
es
out
pe
r
f
or
m
t
w
eet
f
eat
ur
es
and
t
hat
t
he
i
r
u
ni
o
n o
ut
p
er
f
or
m
s
bot
h.
4.
2.
F
e
at
u
r
e
S
el
ect
i
o
n
I
n or
der
t
o
u
nd
er
s
t
and t
he
ov
er
al
l
c
ont
r
i
bu
t
i
o
n of
eac
h i
n
di
v
i
du
al
f
eat
ur
e f
or
pr
e
di
c
t
i
on
ac
c
ur
ac
y
,
w
e u
t
i
l
i
z
e
r
ec
ur
s
i
v
e
f
eat
ur
e
el
i
m
i
nat
i
on (
R
F
E
)
an
d c
om
put
e
gi
ni
i
m
por
t
an
c
e
on
al
l
f
eat
ur
es
i
n t
he g
l
o
ba
l
l
i
ne
ar
and
non
-
l
i
near
m
odel
s
.
R
F
E
s
t
ar
t
s
b
y
t
r
a
i
ni
ng m
ode
l
s
w
i
t
h al
l
f
eat
ur
es
.
T
he f
eat
ur
e
w
hos
e abs
o
l
ut
e
w
ei
ght
i
s
t
hen
f
ound s
m
al
l
es
t
i
s
pr
u
ne
d f
r
om
t
he s
et
;
R
F
E
c
ont
i
n
ues
l
i
k
e t
hi
s
r
ec
ur
s
i
v
el
y
unt
i
l
t
her
e i
s
o
nl
y
o
ne
f
ea
t
ur
e l
ef
t
.
T
hi
s
l
as
t
f
eat
u
r
e i
s
t
he f
i
r
s
t
-
r
ank
ed f
eat
ur
e i
n T
abl
e
6.
Meas
ur
i
ng t
he i
m
por
t
anc
e
of
a f
eat
ur
e i
n d
ec
i
s
i
o
n t
r
ee or
r
an
dom
f
or
es
t
equal
s
c
om
put
i
ng
t
h
e
dec
r
eas
e
of
i
m
pur
i
t
y
of
t
he
nod
es
ov
er
al
l
t
r
e
es
i
n
t
he
f
or
es
t
[
16]
.
T
he l
o
w
er
d
ec
r
eas
i
ng
is
hi
gher
t
h
e
i
m
por
t
anc
e of
t
h
e f
eat
ur
e
i
n t
he
dec
i
s
i
on t
r
ee or
r
an
dom
f
or
es
t
.
T
abl
e
5
.
C
om
par
i
s
on of
F
1
(
%
)
s
c
or
es
f
r
o
m
di
f
f
er
ent
c
l
as
s
i
f
i
er
s
and f
eat
ur
e s
et
s
on
t
he t
es
t
s
et
.
I
n t
h
e r
i
g
ht
m
os
t
c
ol
um
n,
s
t
at
i
s
t
i
c
a
l
l
y
s
i
gn
i
f
i
c
ant
d
i
f
f
er
enc
es
w
i
t
h
t
he
pr
e
v
i
o
us
r
o
w
(
i
n t
h
e s
am
e
par
t
of
t
he
t
ab
l
e)
ar
e m
ar
k
e
d
w
i
t
h
↑
M
ode
l
U
ser
f
eat
u
r
es
T
w
eet
f
eat
u
r
es
Bo
t
h
G
l
obal
m
odel
s
G
-
PA
5
0
.
8
1
3
0
.
2
6
5
1
.
5
8
G
-
LS
V
5
1
.
4
8
1
8
.
9
0
5
4
.
6
5
↑
G
-
LR
5
5
.
1
8
2
3
.
0
1
5
8
.
5
5
↑
G
-
DT
6
8
.
0
1
4
6
.
3
9
6
7
.
9
8
↑
G
-
RF
6
9
.
6
3
3
8
.
7
8
.
↑
T
S
m
odel
s
TS
-
PA
5
1
.
1
5
3
0
.
3
1
5
1
.
8
4
TS
-
L
SV
5
1
.
5
2
1
8
.
9
8
5
4
.
6
6
↑
TS
-
LR
5
5
.
1
8
2
3
.
0
5
5
8
.
5
9
↑
TS
-
DT
6
8
.
0
8
4
6
.
3
9
6
7
.
9
7
↑
TS
-
RF
7
1
.
0
6
4
6
.
3
2
.
↑
T
abl
e
6
.
T
he t
op
-
5
-
f
eat
ur
e r
ank
i
ngs
gen
er
at
e
d f
r
om
R
F
E
f
or
gl
o
ba
l
pas
s
i
v
e
-
a
ggr
es
s
i
v
e (
G
-
P
A)
,
gl
o
bal
l
i
ne
ar
s
upp
or
t
v
ec
t
or
(
G
-
LS
V
)
,
a
nd
gl
o
ba
l
l
og
i
s
t
i
c
r
egr
es
s
i
on (
G
-
LR
)
Ra
n
k
G
-
PA
G
-
LS
V
G
-
LR
1
N
um
ber
o
f
u
s
er
l
i
s
t
ed
N
um
ber
o
f
u
s
er
l
i
s
t
ed
N
um
ber
o
f
u
s
er
l
i
s
t
ed
2
I
s
a u
s
er
v
er
i
f
i
ed?
#f
ol
l
ow
er
s
#f
ol
l
ow
er
s
3
#f
ol
l
ow
er
s
A
v
er
age #t
w
eet
s
/
day
A
v
er
age #t
w
eet
s
/
day
4
A
v
er
age
#t
w
eet
s
/
day
#f
ol
l
ow
er
s
/
#
f
r
i
ends
#f
ol
l
ow
er
s
/
#
f
r
i
ends
5
#f
ol
l
ow
er
s
/
#
f
r
i
ends
Lengt
h of
a
t
w
eet
I
s
a u
s
er
v
er
i
f
i
ed?
#t
w
eet
s
=
nu
m
ber
of
t
w
eet
s
,
#f
ol
l
o
w
er
s
=
nu
m
ber
of
f
ol
l
ow
er
s
,
#
f
r
i
en
ds
=
nu
m
ber
of
f
r
i
end
s
T
abl
e 6
s
ho
w
s
t
hat
t
h
e
nu
m
ber
o
f
t
i
m
es
a
us
er
i
s
l
i
s
t
ed,
t
he
n
um
ber
of
f
ol
l
o
w
er
s
,
and
t
he
av
er
age
num
ber
of
t
w
eet
s
p
ubl
i
s
he
d p
er
da
y
ar
e t
h
e f
eat
ur
es
t
hat
c
ont
r
i
b
ut
e m
os
t
t
o t
h
e
pr
edi
c
t
i
on ac
c
ur
ac
y
.
M
or
e
ov
er
,
t
he
i
m
por
t
ant
f
eat
ur
es
of
our
gl
o
bal
r
and
om
f
or
es
t
m
odel
i
n
T
abl
e 7 ar
e s
i
m
i
l
ar
t
o t
he
o
nes
i
d
ent
i
f
i
ed
b
y
R
F
E
f
or
t
h
e l
i
ne
ar
l
e
ar
ner
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NI
K
A
I
S
S
N
:
1
693
-
6
930
A
C
o
mpar
i
s
on
of
R
e
t
w
ee
t
P
r
ed
i
c
t
i
o
n A
ppr
oac
hes
:
T
h
e
S
u
per
i
or
i
t
y
…
(
H
endr
a
B
u
ny
ami
n
)
1057
T
abl
e
7
.
T
he t
op
-
5
-
f
eat
ur
e
i
m
por
t
anc
e r
ank
i
ngs
f
or
gl
o
bal
dec
i
s
i
on t
r
ee c
l
as
s
i
f
i
er
(
G
-
D
T
)
and
gl
o
bal
r
and
om
f
or
es
t
c
l
as
s
i
f
i
c
at
i
on (
G
-
R
F
)
bas
ed
on
gi
ni
i
m
por
t
anc
e
Ra
n
k
G
-
DT
G
-
RF
1
N
um
ber
o
f
u
s
er
l
i
s
t
ed
N
um
ber
of
us
e
r
l
i
s
t
e
d
2
A
v
er
age t
w
eet
s
per
day
N
um
ber
o
f
f
ol
l
ow
er
s
3
N
ov
el
t
y
A
v
er
age t
w
eet
s
per
day
4
A
v
er
age t
w
eet
s
per
day
N
um
ber
o
f
f
ol
l
ow
er
s
/
N
u
m
ber
of
f
r
i
ends
5
A
c
c
ount
age
N
ov
el
t
y
I
nt
er
es
t
i
ng
l
y
,
b
ot
h
gr
aphs
i
n
F
i
g
ur
e
1
s
ho
w
t
h
at
t
her
e
i
s
m
or
e
t
han
10%
i
nc
r
eas
e
of
F
1
s
c
or
e (
bl
ue l
i
ne)
w
he
n
w
e a
dd t
he f
eat
ur
e,
n
um
ber
of
t
w
eet
s
pub
l
i
s
h
ed.
W
e
f
i
nd t
hat
G
-
LS
V
and
G
-
LR
w
i
t
h s
el
ec
t
i
ng
4 b
es
t
f
eat
ur
es
pl
us
”
num
ber
of
t
w
eet
s
p
ubl
i
s
he
d”
c
an
ac
hi
ev
e per
f
or
m
anc
e
c
o
m
par
abl
e
w
i
t
h
t
he
p
er
f
or
m
anc
e
o
f
bot
h
m
odel
s
ut
i
l
i
z
i
ng
a
l
l
t
h
e
f
e
at
ur
es
(
gr
ee
n
das
hed
l
i
n
e
i
n
F
i
gur
e 1)
.
F
i
gur
e 1.
(
T
op)
:
F
1
sco
r
e
s o
f
G
-
LS
V
t
r
ai
ned
i
nc
r
em
ent
a
l
l
y
w
i
t
h m
or
e f
eat
ur
es
(
x
-
a
x
i
s
)
. (
B
o
tto
m
)
:
F
1
s
c
or
es
of
G
-
LR
.
F
1
s
c
or
es
ar
e c
om
put
ed f
r
o
m
5
-
f
ol
d c
r
os
s
-
v
al
i
d
at
i
on,
t
he
or
der
of
t
he
f
eat
ur
es
i
s
det
er
m
i
ned b
y
R
F
E
c
om
put
ed f
r
o
m
t
he c
or
r
es
pond
i
n
g m
odel
.
T
he d
as
hed
gr
een
l
i
ne
dep
i
c
t
s
t
he F
1
s
c
or
e
of
t
he
c
or
r
es
pondi
ng m
odel
t
r
ai
ne
d on
t
he
bes
t
4
f
eat
ur
es
+
t
he n
um
ber
of
t
w
eet
s
p
ub
l
i
s
he
d (
t
h
e f
eat
u
r
e c
aus
i
n
g t
h
e bo
os
t
i
n F
1
s
c
or
e f
or
bot
h m
odel
s
)
4
.3
. T
im
e
-
sen
s
i
t
i
v
e M
o
d
e
l
i
n
g
T
im
e
-
s
e
n
s
it
iv
e
m
odel
i
n
g g
e
ner
al
l
y
y
ie
ld
s
bet
t
er
per
f
or
m
anc
e t
han t
he
gl
oba
l
m
odel
s
;
t
h
e
ex
c
ept
i
o
n i
s
G
-
D
T
,
w
hi
c
h
out
p
er
f
or
m
s
T
S
-
D
T
b
y
0.
0
1%
.
H
o
w
ev
er
,
t
h
e Mc
N
em
ar
s
i
gn
i
f
i
c
anc
e
t
es
t
do
es
not
i
n
di
c
at
e t
h
at
a
n
y
of
t
he d
i
f
f
er
enc
es
ar
e s
i
gni
f
i
c
ant
.
W
e
al
s
o
c
ar
r
i
ed
out
ex
per
i
m
ent
s
t
o
f
i
nd
t
he
opt
i
m
al
gl
ob
al
λ
i
n
equat
i
o
n
(
3)
and
(
4)
as
t
r
y
i
n
g i
nd
i
v
i
dua
l
λ
f
or
eac
h
l
oc
al
m
odel
i
s
i
nt
r
ac
t
a
bl
e
.
H
o
w
e
v
er
,
t
he
l
oc
al
m
odel
i
ns
i
d
e t
i
m
e
-
s
ens
i
t
i
v
e m
odel
i
n
g
is
s
t
i
l
l
u
nab
l
e t
o c
ont
r
i
but
e t
o
he
l
p
i
m
pr
ov
e t
he pr
edi
c
t
i
o
ns
.
W
e
c
onc
l
ude
t
hat
,
on
our
dat
as
et
an
d
u
nl
i
k
e
t
he
f
i
nd
i
ngs
b
y
P
et
r
ov
i
c
,
et
al
.
,
[
3
],
ti
m
e
-
s
ens
i
t
i
v
e m
odel
i
n
g has
a
v
er
y
l
i
m
i
t
ed c
ont
r
i
but
i
o
n t
o t
he o
v
er
a
l
l
per
f
or
m
anc
e
t
he
us
e of
a s
t
r
on
g
l
ear
n
i
n
g m
odel
i
s
f
ar
m
or
e i
m
por
t
ant
.
5
.
C
o
n
c
l
u
s
i
o
n
W
e
pr
ov
i
d
e a
s
t
ud
y
a
nd c
om
par
i
s
on of
r
et
w
eet
pr
e
di
c
t
i
on
appr
o
ac
hes
.
T
o t
h
e
bes
t
of
our
k
now
l
e
dge
,
t
hi
s
i
s
t
h
e
f
i
r
s
t
at
t
em
pt
t
o c
o
l
l
ec
t
b
es
t
per
f
or
m
i
ng f
eat
ur
es
a
nd c
ont
r
as
t
l
i
nea
r
and non
-
l
i
n
ear
l
ear
n
i
n
g m
et
hods
.
S
pec
i
f
i
c
a
l
l
y
,
w
e
ans
w
er
t
hr
ee r
es
e
ar
c
h ques
t
i
ons
.
W
e
dem
ons
t
r
at
e
t
hat
,
on
our
d
at
as
et
,
a
r
and
om
f
or
es
t
-
bas
ed
l
ear
n
i
ng
m
et
hod,
w
h
i
c
h
has
not
bee
n
em
pl
o
y
ed
i
n
pr
e
v
i
o
us
s
t
ud
i
es
,
out
p
er
f
or
m
s
al
l
ot
h
er
l
e
ar
ni
n
g m
et
hods
t
hat
w
e c
o
ns
i
der
.
W
e
f
i
nd
t
hat
us
er
f
eat
ur
es
ar
e
m
or
e
i
m
por
t
ant
t
h
an t
w
eet
f
eat
ur
es
i
n m
ak
i
ng c
or
r
ec
t
pr
edi
c
t
i
ons
and
t
ha
t
t
he
bes
t
t
hr
ee
f
eat
ur
es
ar
e:
num
ber
of
t
i
m
es
a
us
er
i
s
l
i
s
t
e
d,
n
um
ber
of
f
ol
l
o
w
er
s
,
and
a
v
er
a
ge
num
ber
of
t
w
eet
s
pu
bl
i
s
he
d
per
da
y
.
U
s
i
ng
f
eat
ur
e
s
el
ec
t
i
on,
w
e
f
i
nd
t
h
at
t
h
e
nu
m
ber
o
f
t
w
eet
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SSN
:
1
6
9
3
-
6
930
T
E
L
KO
M
NI
K
A
V
o
l.
14
,
N
o
.
3,
S
ept
em
ber
2016
:
10
52
–
1
058
1058
pub
l
i
s
he
d c
om
bi
ned
w
i
t
h t
h
e
f
our
bes
t
f
eat
ur
es
l
e
ads
t
o per
f
or
m
anc
e l
ev
el
s
of
G
-
LS
V
an
d G
-
LR
m
odel
s
t
hat
ar
e c
om
par
a
bl
e
us
i
ng
al
l
f
eat
ur
es
.
L
a
s
t
l
y
,
t
i
m
e
-
s
ens
i
t
i
v
e m
odel
i
ng h
as
l
i
m
i
t
ed
benef
i
t
s
o
n o
ur
dat
as
et
.
A
s
t
o f
ut
ur
e
w
or
k
,
w
e pl
a
n t
o s
t
ud
y
t
he
pot
e
nt
i
al
o
f
s
i
gnal
s
i
nf
er
r
ed f
r
o
m
ex
t
er
na
l
s
our
c
es
(
s
uc
h as
ne
w
s
or
W
i
k
i
pedi
a)
f
or
r
et
w
e
et
pr
e
di
c
t
i
on.
R
ef
er
en
ces
[1
]
G
oy
al
S
.
F
ac
eboo
k
,
T
w
i
t
t
er
,
G
oogl
e+
:
S
o
c
i
a
l
N
et
w
or
k
i
ng.
I
nt
er
nat
i
on
al
J
o
ur
na
l
of
S
oc
i
a
l
N
et
w
or
k
i
ng
and V
i
r
t
ual
C
om
m
uni
t
i
e
s
.
201
2;
1(
1)
.
[2
]
K
w
ak
H
,
Lee C
,
P
ar
k
H
,
M
oon S
.
W
h
at
i
s
T
w
i
t
t
er
,
a S
oc
i
al
N
et
w
or
k
or
a
N
ew
s
M
edi
a?
P
r
oc
e
e
di
n
gs
of
t
he 19
t
h I
nt
er
n
at
i
o
nal
C
onf
er
e
nc
e
on
W
o
r
l
d
W
i
de
W
e
b.
R
al
e
i
gh,
N
or
t
h
C
ar
ol
i
na,
U
S
A
.
20
10
:
591
-
6
00.
[3
]
Pe
t
ro
v
i
c
S
,
O
s
bor
n
e
M
,
Lav
r
enk
o
V
.
R
T
t
o
W
i
n!
P
r
edi
c
t
in
g
M
es
s
age
P
r
opa
gat
i
on
i
n
T
w
i
t
t
er
.
Pro
c
ee
di
n
gs
of
t
h
e F
i
f
t
h I
nt
er
nat
i
o
nal
C
on
f
er
en
c
e on
W
e
bl
ogs
an
d S
oc
i
al
M
edi
a
.
B
ar
c
e
l
ona,
S
pa
i
n
.
2011.
[4
]
H
ong L,
D
an O
,
D
av
i
s
on B
D
.
P
r
edi
c
t
i
ng P
o
pul
ar
M
es
s
ages
i
n T
w
i
t
t
er
.
I
n
t
e
r
na
t
i
on
a
l
W
or
l
d
W
ide
W
eb
C
onf
er
en
c
e
s
.
H
y
der
a
bad,
I
ndi
a.
201
1:
57
-
58
.
[5
]
N
av
eed N
,
G
ot
t
r
o
n T
,
K
unegi
s
J
,
A
l
had
i
A
C
.
B
ad N
ew
s
T
r
av
el
F
a
s
t
:
A
C
ont
en
t
-
ba
s
ed A
nal
y
s
i
s
of
I
nt
er
e
s
t
i
n
gne
s
s
on
T
w
i
t
t
er
.
W
e
b S
c
i
enc
e C
onf
er
en
c
e.
K
obl
e
n
z
,
G
er
m
any
.
20
11:
8.
[6
]
Art
z
i
Y
,
Pa
n
t
e
l
P,
G
a
m
o
n
M
.
P
r
edi
c
t
i
n
g R
e
s
pon
s
e
s
t
o M
i
c
r
obl
o
g P
o
s
t
s
.
H
um
an
Lang
uag
e
T
ec
hnol
o
gi
e
s
:
C
onf
er
en
c
e
of
t
he
N
or
t
h
A
m
er
i
c
a
n
C
hapt
er
of
t
he
A
s
s
oc
i
at
i
on
of
C
o
m
put
at
i
o
nal
Li
ngu
i
s
t
i
c
s
P
r
o
c
ee
di
n
gs
.
M
ont
r
eal
,
C
a
nada
.
20
12:
602
-
6
06.
[7
]
X
u Z
,
Y
ang Q
.
A
nal
y
z
i
ng U
s
er
R
et
w
e
et
B
eh
av
i
or
on T
w
i
t
t
er
.
T
he I
nt
er
nat
i
on
al
C
on
f
er
en
c
e o
n
A
dv
anc
e
s
i
n
S
oc
i
al
N
et
w
or
k
A
nal
y
s
i
s
an
d M
i
ni
ng.
C
a
l
gar
y
,
C
anada
.
20
12:
46
-
5
0.
[8
]
C
ar
uana
R
,
N
i
c
ul
e
s
c
u
-
M
iz
il
A
.
A
n
E
m
pi
r
i
c
al
C
om
pa
r
i
s
on
of
S
upe
r
v
i
s
ed
Lear
n
i
ng
A
l
gor
i
t
hm
s
.
P
r
oc
ee
di
n
gs
of
t
h
e 23
r
d I
n
t
er
nat
i
o
nal
C
onf
er
enc
e o
n M
ac
hi
ne Le
ar
ni
ng.
P
i
t
t
s
bur
g
h,
U
S
A
.
2006
:
1
61
-
168.
[9
]
F
er
nan
dez
-
D
el
ga
do
M
,
C
er
nadas
E
,
B
ar
r
o S
,
A
m
or
i
m
D
.
D
o w
e need hundr
ed
s
of
c
l
as
s
i
f
i
er
s
t
o s
ol
v
e
r
eal
w
or
l
d c
l
as
s
i
f
i
c
a
t
i
o
n pr
ob
l
em
s
?
T
he
J
our
nal
of
M
ac
hi
ne
Lear
n
i
ng R
e
s
ear
c
h
.
2014
;
1
5(
1)
:
31
33
-
3181.
[
10]
J
end
er
s
M
,
K
as
nec
i
G
,
N
aum
ann
F
.
A
nal
y
z
i
ng
and
P
r
edi
c
t
i
ng
V
i
r
al
T
w
eet
s
.
In
t
e
r
n
a
t
i
o
n
al
W
or
l
d
W
ide
W
e
b
C
on
f
er
en
c
e
s
.
R
i
o
de
J
an
ei
r
o,
B
r
az
i
l
.
2
013:
657
-
664.
[
11]
G
ao S
,
M
a J
,
C
hen Z
.
M
odel
i
ng an
d P
r
edi
c
t
i
ng R
et
w
ee
t
i
n
g D
y
nam
i
c
s
on M
i
c
r
obl
ogg
i
ng
P
l
at
f
or
m
s
.
P
r
oc
ee
di
n
gs
of
t
h
e E
i
ght
h A
C
M
I
nt
er
nat
i
o
nal
C
o
nf
er
e
nc
e o
n
W
eb S
e
ar
c
h and D
at
a M
i
ni
ng
.
N
ew
Y
or
k
C
i
t
y
,
U
S
A
.
2015
:
107
-
116.
[
12]
Z
am
a
n
T
,
F
o
x
E
B
,
B
r
adl
ow
E
T
.
A
B
ay
es
i
an
A
ppr
oa
c
h
f
or
P
r
edi
c
t
i
ng
t
he
P
op
ul
ar
i
t
y
of
T
w
eet
s
.
T
h
e
A
nnal
s
of
A
ppl
i
ed S
t
at
i
s
t
i
c
s
.
2
014;
8(
3)
:
1
583
-
1611
.
[
13]
M
a
cska
s
sy
S
A
,
M
i
ch
e
l
so
n
M
.
W
h
y
d
o P
eo
pl
e R
e
t
w
eet
? A
nt
i
-
hom
ophi
l
y
W
i
n
s
t
he
D
a
y
!
P
r
o
c
eed
i
ng
s
of
t
he F
i
f
t
h I
nt
er
n
at
i
o
nal
C
onf
er
e
nc
e
on
W
e
b
l
o
gs
and S
oc
i
al
M
edi
a.
B
ar
c
el
on
a,
S
p
ai
n
.
201
1.
[
14]
M
or
c
hi
d
M
,
D
uf
our
R
,
B
ous
quet
P
-
M
,
Li
nar
es
G
,
T
or
r
es
-
M
or
eno
J
-
M
.
F
eat
ur
e
S
el
e
c
t
i
o
n
us
i
ng
P
r
i
nc
i
pal
C
om
pone
nt
A
nal
y
s
i
s
f
or
m
as
s
i
v
e r
et
w
eet
de
t
ec
t
i
o
n.
P
at
t
er
n R
e
c
og
ni
t
i
on
Let
t
er
s
.
201
4;
49
:
33
-
39.
[
15]
S
uh B
,
H
ong L,
P
i
r
ol
l
i
P
,
C
hi
E
H
.
W
an
t
t
o be R
et
w
eet
ed? L
ar
ge S
c
a
l
e A
nal
y
t
i
c
s
o
n F
ac
t
o
r
s
I
m
pac
t
i
ng
R
et
w
eet
i
n T
w
i
t
t
er
N
et
w
or
k
.
2010 I
E
E
E
S
ec
ond I
n
t
er
n
a
t
i
ona
l
C
onf
er
enc
e on S
o
c
i
a
l
C
om
put
i
ng
(
S
oc
i
a
l
C
om
)
.
M
i
nneapo
l
i
s
,
U
S
A
.
2010
:
1
77
-
184
.
[
16]
P
edr
ego
s
a
F
,
V
ar
o
qua
ux
G
,
G
r
am
f
or
t
A
,
M
i
c
hel
V
,
T
hi
r
i
on
B
,
G
r
i
s
e
l
O
,
B
l
o
ndel
M
,
P
r
et
t
enhof
er
P
,
W
e
i
s
s
R
,
D
ubo
ur
g V
,
V
an
der
pl
as
J
,
P
a
s
s
o
s
A
,
C
our
nape
a
u D
,
B
r
uc
her
M
,
P
er
r
ot
M
,
D
u
c
he
s
nay
E
.
S
c
ik
it
-
l
e
ar
n:
M
ac
hi
ne
L
ear
ni
n
g
i
n
P
y
t
ho
n.
J
our
n
al
o
f
M
ac
hi
ne
L
ear
n
i
ng
R
e
s
ear
c
h
.
201
1;
12:
2
825
-
2830.
[
17]
W
a
n
g L.
M
ac
hi
n
e Lear
n
i
ng i
n
B
i
g D
at
a.
I
nt
e
r
nat
i
ona
l
J
o
ur
n
al
of
A
dv
anc
es
i
n A
ppl
i
ed S
c
i
enc
e
.
2016
:
4
(4
).
[
18]
B
r
ei
m
a
n L.
R
ando
m
F
or
es
t
s
.
M
ac
hi
ne Le
ar
ni
n
g
.
2
001;
45(
1)
:
5
-
32.
Evaluation Warning : The document was created with Spire.PDF for Python.