Int
ern
at
i
onal
Journ
al of Ele
ctrical
an
d
Co
mput
er
En
gin
eeri
ng
(IJ
E
C
E)
Vo
l.
8
, No
.
6
,
Decem
ber
201
8,
pp. 4
554~
4567
IS
S
N: 20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v8
i
6
.
pp
4554
-
45
67
4554
Journ
al h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
A Novel
Hybrid
Classi
fication
Ap
proach f
or Senti
ment An
alys
i
s
of
T
ext Do
cume
nt
Ya
s
sine
Al Amr
an
i
1
,
M
oha
med Laz
aar
2
,
Ka
m
al Eddin
e El
Kadir
i
3
1,3
LIROSA
La
bo
rat
or
y
,
Abdelma
le
k
Essaa
di
Univ
ersity
,
Morocc
o
2
New T
e
chnol
og
y
Tre
nds
Team
,
Abdelmale
k
Essaa
di
Univer
sit
y
,
Morocc
o
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Feb
1
1
, 201
8
Re
vised
Jun
2
0
, 201
8
Accepte
d
J
ul
22
, 2
01
8
Senti
m
ent
anal
ysis
is
a
m
ore
popula
r
area
of
highly
a
ct
iv
e
rese
arc
h
i
n
Autom
at
ic
La
ng
uage
Proce
ss
ing.
She
assigns
a
n
ega
t
ive
or
positi
ve
polar
i
t
y
to
one
or
m
ore
e
nti
ties
using
diff
ere
nt
n
at
ur
al
la
n
guage
pro
ce
ss
in
g
tool
s
and
al
so pre
di
cted
hi
gh
and
low
p
erf
o
rm
anc
e
of
v
ari
o
us sent
iment
c
las
sifie
rs.
Our
appr
oa
ch
fo
cuse
s
on
the
an
aly
sis
of
fe
el
ings
result
ing
from
rev
ie
ws
of
produc
ts
using
origi
nal
t
ext
sea
rch
techniqu
es.
The
se
rev
iew
s
ca
n
be
cl
assifi
ed
as
h
av
ing
a
positi
v
e
or
negative
fe
el
ing
base
d
on
c
ert
a
i
n
aspe
c
ts
in
rel
a
ti
on
to
a
qu
er
y
base
d
on
term
s.
In
thi
s
pa
per
,
we
chose
t
o
use
two
aut
om
at
i
c
le
arn
i
ng
m
et
hods
for
cl
assificat
ion
:
Support
Vec
tor
Mac
hine
s
(SV
M)
and
Ran
dom
Forest,
and
we
int
roduc
e
a
novel
h
y
br
id
a
pproa
ch
to
ide
nti
f
y
produ
ct
rev
ie
ws
offe
red
b
y
Am
az
on.
T
his
is
useful
for
consum
ers
who
want
to
re
sea
rch
the
se
nti
m
ent
of
pro
duct
s
bef
ore
p
urc
hase
,
o
r
companie
s
th
at
want
to
m
onit
or
the
publ
ic
sen
t
iment
of
th
ei
r
b
ran
ds.
Th
e
result
s
summ
ari
ze
tha
t
th
e
proposed
m
et
hod
o
utpe
rform
s
the
se
indi
vidual
cl
assifi
ers
in
thi
s
amaz
on
dataset
.
Ke
yw
or
d:
Am
azon
Cl
assifi
ers
Ra
ndom
Fo
rest
Sentim
ent A
na
ly
sis
Suppor
t
V
ect
or Mac
hin
e
Copyright
©
201
8
Instit
ut
e
o
f Ad
vanc
ed
Engi
n
ee
r
ing
and
S
cienc
e
.
Al
l
rights re
serv
ed
.
Corres
pond
in
g
Aut
h
or
:
Yassine
A
l
A
m
ran
i,
LIROSA
La
bora
tor
y
,
Abdelm
al
ek
Esaadi U
niv
e
rsity
,
Tet
uan, Mo
r
oc
co
.
Em
a
il
:
al
a
m
ran
iy
assine@g
m
ai
l.com
1.
INTROD
U
CTION
Cl
assifi
cat
ion
is
the
process
wh
e
rein
a
cl
as
s
la
bel
is
assigned
to
unla
bel
ed
data
vect
ors.
It
can
be
cat
egorized
int
o
supe
rv
ise
d
a
nd
un
-
s
uper
vis
ed
cl
assifi
cat
io
n
w
hich
is
al
s
o
known
as
cl
ust
ering.
I
n
s
up
e
r
vised
cl
assifi
cat
ion
l
earn
i
ng
is
do
ne
with
t
he
help
of
s
up
e
rv
is
or
i.e.
le
ar
ning
th
rou
gh
e
xam
ple.
I
n
this
m
et
ho
d,
t
he
set
of
possi
ble
cl
ass
la
bels
is
known
a
pr
io
ri
to
the
en
d
us
e
r
[1]
.
S
uper
vis
ed
cl
assifi
cat
io
n
can
be
sub
div
ide
d
into
no
n
-
pa
ra
m
et
ric
and
par
am
et
ric
cl
as
sific
at
ion
.
Pa
r
a
m
et
ric
cl
assif
ie
r
m
e
tho
d
i
s
dep
e
nd
e
nt
on
the
pro
bab
il
it
y
distribu
ti
on
of
each
cl
ass.
N
on
-
par
am
et
ric
c
la
ssifie
rs
are
us
e
d
wh
e
n
th
e
den
sit
y
func
ti
on
is
unknow
n.
Exa
m
ples
of
par
a
m
et
ric
su
pervised
cl
assifi
cat
ion
m
et
ho
ds
are
Mi
ni
m
al
Dista
nce
Cl
assifi
er
,
Ba
ye
sia
n,
Mul
ti
var
ia
te
Ga
us
s
ia
n,
S
uppo
rt
V
ect
or
m
achines
an
d
Decisi
on
Tree.
Exam
ples
of
non
-
pa
ra
m
et
ric
su
pe
r
vised
cl
assifi
cat
ion
m
et
ho
ds
are
K
-
Nea
rest
Nei
ghbors
,
Eucli
dea
n
Dis
ta
nce,
L
ogist
ic
Re
gr
e
ssio
n,
Neural
Netw
ork Ker
ne
l Densi
ty
Esti
m
at
ion
, Arti
fic
ia
l Neural
N
et
work a
nd Mult
il
ay
er P
erce
ptr
on.
Re
centl
y,
m
ult
iple
platf
or
m
s
are
dev
el
op
i
ng
ver
y
i
nteresti
ng
ei
ther
in
te
rm
s
of
volum
e
of
data
or
accor
ding
to
t
he
num
ber
of
us
ers
a
r
ound
the
w
or
l
d,
th
ey
offer
us
er
s
al
l
th
e
po
s
sibil
it
ie
s
to
ex
pr
ess
their
op
i
nions
an
d
t
o
exc
ha
ng
e
th
ei
r
ideas
with
the
oth
e
rs
[2]
.
The
sentim
ent
analy
sis
f
ound
in
the
f
or
m
of
com
m
ents,
re
views
an
d
fee
db
ac
k
a
nd
prov
i
des
necess
ary
inf
orm
ation
for
var
i
ou
s
pur
poses
[3]
.
The
s
e
op
i
nions
or
senti
m
ents
can
be
div
i
ded
int
o
two
cat
eg
ori
es:
po
sit
ive
and
neg
at
ive;
or
al
so
cat
egories
of
diff
e
re
nt
rati
ng
po
i
nts
(e
.g.
3
sta
rs,
4
sta
rs
a
nd
5
sta
rs).
T
he
po
la
rity
of
se
nti
m
ents
li
ke
“goo
d”
a
nd
“
ba
d”
al
s
o
identify
the
se
nti
m
ents
ei
ther
po
sit
ive
or
ne
gative
[
4]
.
Sen
tim
ent
analy
sis
is
the
par
t
of
the
te
xt
m
ining
that
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel
Hybri
d
Cl
as
sif
ic
atio
n
A
pproac
h
fo
r
S
e
ntiment
…
(
Ya
ssi
ne
Al A
m
ra
ni)
4555
at
tem
pts to
defi
ne
the
op
i
nion
s,
feeli
ngs a
nd
at
ti
tud
es prese
nt in
a text
or
a
set of
te
xt.
It i
s p
arti
cula
rly
used in
m
ark
et
ing
t
o
a
naly
se
for
e
xa
m
ple
the
com
m
ents
of
th
e
Net
surfe
rs
or
the
com
par
at
ives
a
nd
te
sts
of
t
he
blogg
e
rs.
It
req
ui
res
m
uch
m
or
e
underst
an
di
ng
of
the
la
ng
uag
e
tha
n
te
xt
analy
sis
and
s
ubj
ect
cl
assifi
cat
ion
.
Indee
d,
if
the
si
m
plest
al
go
ri
thm
s
con
side
r
on
ly
the
sta
ti
sti
cs
of
f
reque
nc
y
of
occ
urre
nc
e
of
the
w
ords
,
it
is
us
ua
ll
y
insuffi
ci
en
t
to
def
i
ne
the
do
m
inant
opinio
n
i
n
a
do
c
um
ent.
It
i
s
the
process
of
determ
ining
the
con
te
xtu
al
pola
rity
o
f
t
he
te
xt,
that is,
w
hethe
r
a te
xt is posit
ive or
ne
gative
[
5]
.
The
us
e
of
t
his
analy
sis
help
s
researc
he
rs
a
nd
decisi
on
-
m
a
ker
s
bette
r
unde
rstan
d
op
i
nions
a
nd
cl
ie
nt
sat
isfact
ion
us
i
ng
sentim
ent
cl
assifi
cat
ion
te
chn
i
qu
e
s
in
order
t
o
a
uto
m
atical
ly
colle
ct
diff
ere
nt
pe
rs
pe
ct
ives
on
from
var
io
us
platfo
rm
s.
Ther
e
has
been
a
la
rg
e
am
ou
nt
of
researc
h
in
the
area
of
se
nt
i
m
ent
cl
assifi
c
at
ion
.
Trad
it
io
nally
m
os
t
of
it
has
f
ocused
on
cl
as
sifyi
ng
la
r
ge
r
pieces
of
te
xt,
li
ke
rev
ie
ws
(
B.
Pang,
L.
Le
e,
an
d
S
.
Vait
hyanath
an
.
,
20
02).
I
n
this
pa
per,
a
c
om
par
iso
n
of
po
pula
r
cl
assifi
ers
was
perf
or
m
ed
to
cl
assify
product
rev
ie
ws
ei
ther
posit
ive
or
ne
gativ
e:
S
upport
Vector
Ma
chine,
Ra
ndom
Fo
rest
a
nd
our
a
ppro
ac
h
Ra
ndom
Fo
r
est
S
upport
V
ect
or Mac
hin
e
(RFS
VM).
This
pa
per
pre
sents
a
m
et
ho
d
to
determ
ine
how
se
nti
m
ents
can
be
cl
assi
fied
usi
ng
hybri
d
ap
proac
h
of
Sup
port
Ve
ct
or
Ma
c
hin
e
and
Ra
ndom
Fo
r
es
t
.
T
he
pa
per
pro
vid
es
the
c
om
pa
rison
with
oth
e
r
e
xi
sti
ng
te
chn
iq
ue,
s
hows
that
t
he
use
of
hy
br
i
d
a
ppr
oac
h
ca
n
im
pro
ve
the
ef
fici
ency
of
se
nti
m
ent
analy
sis.
T
he
pro
po
se
d
hybri
d
a
pproach
gi
ve
s
bette
r
res
ult
as
com
par
e
t
o
the
e
xisti
ng
te
chn
i
qu
e
s
.
T
he
rest
of
the
pa
pe
r
is
descr
i
bed
as
f
ol
lo
ws:
Sect
io
n
2
desc
ribe
se
nt
i
m
ent
analy
sis
syst
e
m
.
Sect
ion
3
int
rod
uces
app
li
ed
al
gorithm
s
in
this
fiel
d.
Sect
ion
4
discuss
es
pro
pose
d
m
et
ho
ds
.
Se
ct
ion
5
ex
plain
the
res
ults
and
a
naly
sis
ob
ta
ined.
Sect
ion
6
prese
nts the
conclus
ion
a
nd
fu
t
ur
e
work f
or the
pr
opos
e
d wor
k.
2.
SENTIME
NT
ANA
L
YS
I
S SYSTE
M
To
kn
ow
the
opinio
n
of
the
ot
her
pe
ople
wa
s
al
ways
an
im
po
rtant
infor
m
at
ion
el
e
m
ent
du
ri
ng
the
decisi
on
pr
oce
ss.
Be
f
or
e
m
akin
g
decisi
ons
,
pe
op
le
a
re
in
te
rested
e
norm
ou
sly
in
t
he
opinio
ns
of
the
oth
e
r
people
i
n
diff
e
ren
t
a
reas.
T
he
y
consult
the
opinio
ns
of
t
he
oth
e
r
c
on
s
um
ers
befor
e
m
aking
a purc
hase,
o
r
lo
ok
at
the o
pin
i
ons
o
f
the
oth
e
r
pe
op
le
befor
e see
ing
a film
w
it
h
the cinem
a o
r
befor
e
buyi
ng
a d
isc
. Th
a
nks
to the
i
ntern
et
we
ca
n
disc
over
the
op
i
nions
a
nd
t
he
ex
pe
rim
ents
of
ver
y
a
la
r
ge
nu
m
ber
of
pe
op
le
who
a
re
neither
our
fr
ie
nds,
nor
the
ex
per
ts
of
fiel
ds,
but
of
pe
ople
who
can
hav
e
the
sam
e
ta
ste
s
that
us
,
and
th
us
their
op
i
nions
ca
n
be
ver
y
us
e
fu
l
f
or
us
be
fore
m
akin
g
our
c
ho
i
ce
an
d
to
ha
ve
our
own
i
dea
on
a
gi
ve
n
s
ubj
ect
.
To
day,
m
or
e
and
m
or
e
pe
ople
are
giv
i
ng
t
he
ir
op
i
nion
on
diff
e
ren
t
to
pi
cs,
these
opini
on
s
a
re
avail
a
ble
to
ever
y
on
e
on t
he
inter
net.
Accor
ding
to
t
he
s
urveys
[
6]
,
81%
of
the
use
rs
of
the
inter
net
m
ade
at
least
on
ce
t
he
onli
ne
sea
rch
on
a
pro
duc
t
and
a
pproxim
ately
80
%
of
the
m
declare
that
oth
e
r
pe
op
le
ha
ve
a
sign
ific
a
nt
influ
e
nce
on
their
decisi
on
of
pur
chase,
wh
ic
h
r
epr
ese
nts
one
ver
y
a
la
rg
e
num
ber
of
pe
op
le
.
Appro
xim
a
t
el
y
30
%
pro
vid
ed
a
n
op
i
nion
on
a
pro
duct
,
on
a
se
rv
ic
e
or
on
a
pe
rson
on
li
ne
vi
a
a
m
ark
in
g
syst
e
m
,
wh
ic
h
is
no
t
un
im
po
rta
nt
li
ke
nu
m
ber
.
F
or
th
is
reason,
i.e.
than
ks
to
the
in
te
rest
wh
ic
h
th
e
us
ers
s
how
f
or
the
op
i
nions
on
the
product
s
an
d
the
ser
vices,
the
s
uppliers
of
the
arti
cl
es
s
how
ve
ry
a
gr
e
at
at
te
ntion
with
the
de
velo
pm
ent
of
t
he
m
ark
i
ng
syst
e
m
s
[H
off
m
an
(20
08)].
W
it
h
the
e
xp
l
osi
on
of
platf
orm
s
li
ke
the
blogs,
of
t
he
disc
ussi
on
for
um
s,
Peer
-
to
-
Peer
net
wor
k,
and
var
i
ou
s
oth
e
r
ty
pes
of
so
ci
al
m
edia,
the
consum
ers
hav
e
at
thei
r
dis
po
sal
a
pl
at
fo
rm
without
prece
den
t,
of
ra
nge
a
nd
po
wer
,
m
aking
it
po
s
sible
to
sh
are
their
ex
per
im
e
nts
and
to
m
a
rk
their
op
i
nion
(
posit
ive
or
ne
gativ
e)
on
any
pr
oduct
or
se
rv
ic
e.
The
c
om
pan
ie
s
can
m
eet
the
needs
f
or
the
consum
ers
by
carryin
g
ou
t
m
on
it
ori
ng
a
nd
analy
sis
of
the
opinio
ns
t
o
im
pr
ov
e
thei
r
pro
duct
.
S
uch
a
syst
e
m
will
hav
e
fir
stl
y
to
colle
ct
op
inio
ns
of
the
con
s
um
ers
an
d
us
e
rs
in
do
c
um
ents
wh
ic
h
sh
ow
the
s
ubje
ct
ive
op
i
nions
a
nd
s
entences
.
S
ome
tim
es,
that
is
relat
ively
easy
,
as
in
the
cases
of
gr
eat
sit
es
wh
e
re
t
he
op
i
ni
on
s
of
the users
are w
e
ll
stru
ct
ured
s
uch as
for
e
xa
m
ple A
m
azon
.
com
.
Sentim
ent
is
a
visio
n
based
on
em
otion
rath
er
tha
n
reas
on.
It
is
a
ki
nd
of
su
bject
ive
im
p
ressio
n,
not
facts,
al
so
cal
le
d
the
ex
pr
essi
on
of
sen
sit
ive
feeli
ng
in
art
and
li
te
ratur
e
.
Sentim
ent
An
al
ysi
s
is
al
so
a
t
ask
of
natu
ral
la
ngua
ge
pro
cessi
ng
and
i
nfor
m
at
ion
ext
racti
on
t
ha
t
aim
s
to
get
the
feeli
ngs
of
the
wr
it
er
e
xp
resse
d
by
posit
ive
or
neg
at
ive
c
omm
ents,
quest
io
ns
an
d
r
eq
ues
ts,
by
analy
zi
ng
a
gr
eat
num
ber
of
doc
um
ents.
Sentim
ent
analy
sis
is
the
com
pu
ta
ti
onal
te
ch
nique
f
or
extra
ct
ing
,
cl
assify
i
ng,
un
der
sta
nd
ing
a
nd
determ
ining
op
i
nions
e
xpre
ssed
in
va
rio
us
con
te
nt.
It
fo
c
us
es
on
i
den
ti
f
yi
ng
the
opinio
n
or
se
ntim
ent
that
is
held
a
bout
a
n
obj
ect
.
It
us
es
natu
ral
la
ngua
ge
pr
ocessin
g
an
d
c
om
pu
ta
ti
on
al
te
ch
niq
ue
s
to
a
uto
m
at
e
the
extract
ion
or
cl
assifi
cat
ion
of se
nti
m
ent f
r
om
g
ener
al
ly
u
nst
ru
ct
ure
d
te
xt
[7]
.
In
ge
ner
al
,
se
nti
m
ent
analy
sis
aim
s
to
determ
ine
the
sta
te
of
m
ind
of
a
sp
ea
ker
or
a
wr
it
er
with
resp
ect
to
a
subj
ect
or
t
he
ov
erall
ton
e
of
a
do
c
um
ent.
Word
of
m
ou
th
is
the
proces
s
of
passing
in
f
or
m
at
io
n
from
per
son
to
ano
t
her
a
nd
pl
ay
s
an
i
m
po
rta
nt
ro
le
in
cl
ie
nt
s'
decisi
on
m
a
king
ab
out
ser
vices
or
pro
du
ct
s.
In
bu
si
ness
sit
uat
ion
s
,
Wo
r
d
of
m
ou
th
involve
s
consum
ers
w
ho
s
ha
re
at
ti
tud
es,
op
i
nions,
pro
du
ct
s,
or
se
rv
i
ces
with
oth
e
rs.
W
ord of m
ou
th c
omm
un
ic
at
ion
functi
ons
base
d on soc
ia
l net
work
i
ng
[
8]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4554
-
4
567
4556
In
recent
ye
ars
,
the
m
assive
i
ncr
ease
i
n
the
us
e
of
inte
rn
et
and
t
he
exc
ha
ng
e
of
public
op
i
nion
ar
e
the
eng
i
nes
of
senti
m
ent
analy
sis
tod
ay
.
The
W
e
b
is
an
im
m
ense
rep
os
i
tory
of
str
uctu
red
an
d
unstr
uc
ture
d
data.
An
al
yz
in
g
this
data
t
o
e
xtract
la
te
nt
public
opinio
n
an
d
se
ntim
ent
is
a
dif
ficult
ta
sk.
Sentim
ent
ana
ly
sis
can
be usef
ul i
n on
li
ne
pr
oduc
t rev
ie
ws,
rec
omm
end
at
io
ns
,
b
lo
gs,
us
er'
s vi
ews of p
olit
ic
al
cand
i
dates.
3.
APPLIE
D
AL
GORIT
HM
S
To
eval
uate
th
e
perform
ance
of
our
ap
proac
h,
w
e
ch
os
e
to
us
e
tw
o
super
vised
le
ar
ning
al
gorithm
s:
the
rand
om
fo
r
est
al
go
rithm
wh
ic
h
is
a
cl
as
sific
at
ion
al
gor
it
h
m
that
red
uc
es
the
var
ia
nce
of
the
f
or
ecast
s
of
a
decisi
on
tree
a
lon
e,
th
us
im
pr
ovin
g
t
heir
pe
rfor
m
ances,
a
nd
the
Al
gorith
m
of
Sup
port
Vecto
rs
Ma
c
hin
es
or
Larg
e
Ma
rg
i
na
l
Separ
at
or
s
w
hich
is
a
bi
nary
cl
assifi
cat
ion
m
et
ho
d
by
super
vised
le
a
rn
i
ng.
T
hese
ha
ve
bee
n
chosen
beca
use
they
a
re
t
he
m
achine
le
ar
ning
al
gorithm
s
that
of
te
n
gi
ve
t
he
best
re
su
lt
s
f
or
a
utom
at
i
c
cl
assifi
cat
ion
of texts.
Co
ntr
ol f
lo
w of
the
sy
stem
as shown
in Figu
re
1.
Figure
1
.
Co
ntr
ol f
l
ow of the
s
yst
e
m
3.1.
R
andom
Fores
t
Ra
ndom
fo
rest,
wh
ic
h
we
re
f
or
m
al
ly
pr
op
ose
d
in
20
01
by
Leo
Brei
m
an
and
Ad
èl
e
Cut
le
r,
are
pa
rt
of
the
aut
om
a
ti
c
le
arn
ing
te
chn
i
qu
e
s.
This
al
go
rithm
com
bin
es
the
con
cepts
of
ra
ndom
su
bspace
s
and
"baggin
g".
T
he
decisi
on
tree
fo
rest
al
gorit
hm
trai
ns
on
m
ul
ti
ple
decisio
n
trees
dr
ive
n
on
sli
gh
tl
y
diff
e
re
nt
su
bse
ts o
f data
.
Pict
or
ia
l r
ep
re
sentat
i
on of
ra
ndom
f
or
e
st as
sh
ow
n
in
Fi
gur
e 2
.
The
rand
om
fo
rest
is
par
t
of
the
fam
ily
set
m
et
ho
ds
t
ha
t
ta
ke
the
de
ci
sion
tree
as
an
in
div
i
du
al
pr
e
dictor
,
they
are
ba
sed
on
the
m
et
ho
ds
of
ba
ggin
g,
rando
m
iz
ing
outp
uts
an
d
rand
om
su
bspace
e
xc
us
in
g
boos
ti
ng.
This
al
gorithm
is
on
e
of
t
he
best
a
m
on
g
cl
assifi
cat
ion
al
gorit
hm
s
-
able
to
cl
assify
la
rg
e
am
ounts
of
data
with
accu
racy.
It
is
an
ensem
ble
le
ar
ning
m
e
tho
d
f
or
cl
assifi
cat
io
n
an
d
regressi
on
that
co
ns
tr
ucts
a
nu
m
ber
of
dec
isi
on
trees
at
trai
ning
tim
e
a
nd
deliver
s
th
e
cl
ass
that
is
the
m
od
e
of
the
cl
asses
outpu
t
by
ind
ivi
du
al
t
ree
s.
In
ra
ndom
for
est
cl
assifi
cat
ion
m
et
ho
d,
m
any
cl
assifi
ers
are
gen
e
rated
f
ro
m
s
m
al
le
r
subsets
of
the
input
data
an
d
la
te
r
their
in
di
vid
ual
res
ults
are
a
ggre
gate
d
base
d
on
a
vo
ti
ng
m
echa
nis
m
to
ge
ner
at
e
th
e
desire
d
outp
ut
of
the
in
put
data
set
.
This
ensem
ble
le
arn
in
g
str
at
egy
has
rece
ntly
be
com
e
ver
y
popula
r
.
Be
fore
RF,
boos
ti
ng
a
nd
ba
gg
i
ng
wer
e
t
he
only
two
ensem
ble
le
ar
ning
m
e
tho
ds
us
ed.
RF
ha
s
been
extensi
vely
ap
plied
in
var
i
ous
area
s
inclu
ding
m
od
er
n
dru
g
disc
ov
e
r
y,
netw
ork
int
ru
si
on
detect
io
n,
la
nd
cov
e
r
a
naly
sis,
cre
dit rati
ng a
naly
sis, r
em
ote sen
si
ng and
ge
ne
m
ic
ro
arr
ay
s
d
at
a a
naly
sis e
tc
...
[9]
Ther
e
are
tw
o
ways
to
evalua
te
the
err
or
rat
e.
On
e
is
to
spl
it
the
dataset
i
nto
trai
ni
ng
pa
rt
and
te
st
par
t.
W
e
ca
n
em
plo
y
the
train
in
g
par
t
to
bu
il
d
the
fo
rest,
a
nd
the
n
us
e
the
te
st
par
t
to
calcu
la
te
the
err
or
rate.
Anothe
r
way
i
s
to
us
e
the
O
ut
of
Ba
g
(
O
OB)
er
r
or
est
im
a
t
e.
Be
cau
se
rando
m
forests
al
gorithm
cal
culat
es
the
OO
B
er
r
or
du
r
ing
th
e
trai
ni
ng
phase,
we
do
no
t
nee
d
to
spl
it
the
trai
nin
g
data.
Ra
ndom
forest
is
ensem
ble
of
decisi
on trees
, whic
h
a
re
base
d on in
form
at
i
on g
ai
n,
t
he
c
om
pu
ta
ti
on
form
ula is p
resen
t
ed
as:
(
)
=
−
∑
=
1
log
2
(
)
=
(
)
−
(
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel
Hybri
d
Cl
as
sif
ic
atio
n
A
pproac
h
fo
r
S
e
ntiment
…
(
Ya
ssi
ne
Al A
m
ra
ni)
4557
The
ste
p of ra
ndom
f
orest
ca
n be
re
pr
ese
nted
as:
a.
Use bootst
ra
p
t
o
e
xtract
k
sam
ples fr
om
the original tr
ai
ning
sets with
N sa
m
ples f
or
k
ti
m
es,
b.
Esta
blish
k dec
isi
on
tree
s,
c.
Vo
te
acco
r
ding
to
the
cl
assifi
cat
ion
res
ults
of
al
l
decisi
on
trees,
the
vo
ti
ng
res
ul
ts
cal
le
d
confide
nce
scor
e
can
b
e
d
e
scribe
d
as:
=
(
)
(
)
⁄
Figure
2
.
Pict
ori
al
r
ep
rese
ntati
on
of r
a
ndom
f
orest
3.
2
.
Su
pp
ort
Vector
M
achi
ne
The
SV
M
m
e
thod
was
intr
oduce
d
by
Joa
chim
s
[10]
,
then
us
e
d
by
Dr
uc
ke
r
[11]
,
Tai
ra
and
Haru
no
[
12
]
,
a
nd
Ya
ng
a
nd
Liu
[
13]
.
T
he
geo
m
et
ric
SVM
m
et
ho
d
ca
n
be
c
onside
red
as
the
at
te
m
pt
to
fi
nd
,
a
m
on
g
al
l
the
su
r
faces
1
,
2
,
...
of
a
sp
ace
of
dim
ension
s
|T|
wh
ic
h
se
pa
rates
the
posit
ive
le
arn
i
ng
e
xam
ples
from
the
neg
at
ives.
T
he
le
ar
ni
ng
set
is
giv
e
n
by
a
set
of
ve
ct
or
s
as
so
ci
at
ed
wit
h
their
c
la
ss
of
m
e
m
ber
sh
i
p:
(
1
,
1
)
,
(
2
,
2
)
,
…
,
(
,
)
,
,
{
+
1
,
−
1
}
with:
a.
represe
nts the c
la
ss of m
e
m
b
ersh
i
p.
I
n
a t
w
o
-
cl
ass
proble
m
the f
irst cl
as
s corres
ponds
to
a
posit
ive
answer
(
=
+
1
)
, a
nd th
e seco
nd class
corres
ponds to
a n
e
gative a
nswer
(
=
−
1
)
.
b.
re
pr
ese
nts the
vecto
r of
t
he
te
xt num
ber
j
of
the traini
ng set
.
The
S
upport
Vecto
r
Ma
chine
m
e
tho
d
se
pa
rates
the
posit
ive
cl
ass
vector
s
f
r
om
the
neg
at
ive
cl
as
s
vecto
rs by a
hyperplane
d
e
fin
ed by t
he f
ollo
wing e
qu
at
io
n:
⊗
+
,
,
Fo
r
tw
o
cl
asse
s
of
exam
ples
giv
e
n,
the
goa
l
of
SV
M
is
to
fin
d
a
cl
assi
fier
that
will
separ
at
e
the
dat
a
and
m
axi
m
iz
e
the
distance
betwe
en
thes
e
tw
o
c
la
sses
[
14
]
.
Wi
th
S
VM,
this
c
la
ssifie
r
is
a
li
near
cl
as
sifie
r
cal
le
d
hype
r
pla
n
(
)
.
I
n
the
f
ollo
wing
diag
ram
,
we
de
te
rm
ine
a
hype
rp
la
ne
that
se
par
at
es
t
he
tw
o
set
s
of
point
s.
Sepa
rati
on of t
wo sets
with se
par
at
or
a
s s
hown in Fi
gure
3.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4554
-
4
567
4558
Figure
3
.
Se
parat
ion
of tw
o
se
ts wit
h
se
pa
rator
In
ge
ner
al
,
suc
h
a
hype
rp
la
ne
is
no
t
uniqu
e
[
15
]
.
The
SV
M
m
et
hod
determ
ines
the
opti
m
a
l
hype
rp
la
ne by m
axi
m
iz
ing
the m
arg
in: t
he m
arg
in is t
h
e
di
sta
nce b
et
wee
n
the
posit
ive l
abeled
vecto
rs and t
he
neg
at
ive
la
bel
ed
vecto
rs.
T
he
le
ar
ning
s
e
t
is
not
neces
saril
y
li
near
ly
sepa
rab
le
,
va
riables
of
ga
p
are
introd
uced
f
or
al
l
the
.
Th
e
se
ta
ke
i
nto
account
t
he
e
rror
of
cl
assif
ic
at
ion
,
a
nd
m
us
t
sat
isfy
the
fo
ll
owin
g
i
nequali
ti
es:
⊗
+
≥
1
−
⊗
+
≤
1
+
We
ha
ve
t
o
m
i
nim
iz
e
the
fo
ll
ow
i
ng
f
unct
io
n
of
obj
ect
ive
by
ta
ki
ng
int
o
account
t
hese
const
raints:
1
2
‖
‖
2
+
∑
=
1
.
The
first
te
rm
of
t
his
f
un
ct
io
n
co
rr
e
spo
nd
s
t
o
the
siz
e
of
t
he
m
arg
in
a
nd
t
he
sec
ond
te
rm
represe
nts
the
cl
assifi
cat
ion
e
rror,
wh
e
re
re
pr
ese
nts
t
he
num
ber
of
vect
or
s
of
t
he
trai
ni
ng
set
.
Fi
nd
i
ng
the
pr
e
vious
ob
j
e
ct
ive
functi
on
a
m
ou
nts
to
so
lvi
ng
the
f
ol
lowing
quad
r
at
ic
pr
oble
m
:
fin
ding
the
de
ci
sion
functi
on
ℎ
su
c
h t
hat:
ℎ
(
)
=
(
(
)
)
w
her
e:
(
)
=
∑
⊗
+
=
1
(
)
re
pr
ese
nt the
foll
ow
i
ng fu
nction:
a.
if
>
0
then
(
)
=
1
b.
if
<
0
then
(
)
=
−
1
c.
if
=
0
then
(
)
=
0
re
pr
ese
nt the
c
la
ss of m
e
m
ber
sh
ip
,
re
pr
ese
nt the
param
et
ers
to b
e
foun
d
⊗
re
pr
ese
nt the
s
cal
ar prod
uct
of the
v
ect
or
Xi
with the
v
ect
or X
.
The
near
est
po
ints,
w
hich
al
one
are
us
e
d
for
determ
ining
t
he
hy
perplane
,
are
cal
le
d
s
up
port
vect
or
s
.
Hy
pe
rp
la
ne
of
support
vecto
r
m
achine
as
s
how
n
in
Fig
ure
4.
It
is
ob
vious
that
t
her
e
is
a
m
ult
it
ud
e
of
valid
hype
rp
la
ne
but
the
re
m
ark
abl
e
pr
ope
rty
of
the
SV
M
is
that
this
hyper
pla
ne
m
us
t
be
op
tim
a
l
[16]
.
W
e
are
go
i
ng
t
o
lo
ok
for
it
thu
s
m
ore
a
m
on
g
the
va
li
d
hype
rp
la
ne
s,
the
on
e
w
ho
cr
os
ses
"i
n
t
he
m
idd
le
"
po
i
nts
of
bo
t
h
cl
asses
of
exam
ples.
I
nt
uiti
vely
,
it
co
m
es
dow
n
to
l
ooking
f
or
the
"safest"
hype
r
plane.
I
ndeed
,
le
t
us
su
pp
os
e
t
hat
a
n
e
xam
ple
was
not
pe
rfec
tl
y
descr
i
bed,
a
s
m
al
l
var
ia
ti
on
will
no
t
m
od
ify
it
s
cl
assifi
cat
ion
i
f
it
s
distance
in
th
e
hyperplane
i
s
big
.
F
orm
all
y,
this
a
m
ou
nt
s
to
lookin
g
f
or
a
hyperpla
ne
w
hose
m
in
i
m
u
m
distance
to
the
le
arn
in
g
exam
ples
is
m
axi
m
um
.
This
distance
is
cal
le
d
"m
arg
in"
bet
we
en
the
hype
r
plane
an
d
the ex
am
ples.
The o
pti
m
al
separ
at
or
hype
r
plane
is t
he on
e
that m
axi
m
iz
es
the m
arg
in
[17]
.
In
t
uiti
vely
,
the
fact
of
hav
i
ng
a
wi
der
m
arg
i
n
gets
m
or
e
se
cur
it
y
w
he
n
cl
assify
ing
a
ne
w
e
xam
ple.
More
ov
e
r,
if
w
e
find
the
cl
ass
ifie
r
wh
ic
h
be
ha
ves
best
wit
h
resp
ect
to
the
l
earn
i
ng
data,
it
i
s
cl
ear
that
i
t
will
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel
Hybri
d
Cl
as
sif
ic
atio
n
A
pproac
h
fo
r
S
e
ntiment
…
(
Ya
ssi
ne
Al A
m
ra
ni)
4559
al
so
be
the
one
w
ho
will
at
be
st
al
low
to
cl
assify
the
ne
w
e
xam
ples.
On
the
on
e
ha
nd
F
i
gure
5
s
hows
us
that
with
an
opti
m
a
l
hyper
pla
ne,
a
new
exam
ple
rem
ai
ns
cl
assif
ie
d
well
wh
il
e
it
fall
s
in
the
m
arg
in.
O
n
the
oth
er
hand, we
noti
ce o
n
the
F
ig
ure
6
t
hat w
it
h a s
m
al
le
r
m
arg
in, t
he
exam
ple se
es it
sel
f
ba
dly
cl
assifi
ed.
Figure
4
.
Hy
pe
rp
la
ne of s
up
port
vecto
r
m
achine
Figure
5
.
Be
st
hype
rp
la
ne
se
pa
rator
Figure
6
.
Hy
pe
rp
la
ne wit
h
lo
w
m
arg
in
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4554
-
4
567
4560
In
ge
ner
al
,
t
he
cl
assifi
cat
ion
of
a
ne
w
un
know
n
exam
ple
is
giv
e
n
by
it
s
posit
ion
rela
ti
ve
to
the
op
ti
m
al
h
yperp
la
ne.
f
or
e
xa
m
ple,
in
the
F
igure
5,
the
ne
w
el
em
ent
will
be
cl
assifi
ed
in
the
cat
e
gory
of
re
d
balls i
ns
te
a
d of g
reen ball
s.
4.
PROP
OSE
D MET
HO
D
In
t
his
arti
cl
e,
we
pro
po
se
a
m
e
tho
d
w
hic
h
c
om
bin
es
th
e
powe
r
a
nd
the
ca
pab
il
it
ie
s
of
Ra
ndom
Fo
r
est
an
d
S
uppo
rt
Vect
or
Ma
chines
at
the
sam
e
tim
e
for
the
s
up
e
r
vi
sed
ta
s
ks
to
s
olv
e
t
he
pro
bl
e
m
of
cl
assifi
cat
ion
.
Firstl
y,
Ra
ndom
fo
rest
is
an
ensem
ble
le
ar
ning
m
et
ho
d
that
co
ns
tr
uct
a
nu
m
ber
of
de
ci
sion
trees
at
ran
do
m
ly
sel
ect
ed
featur
es
an
d
pr
edict
the
cl
ass
of
a
te
st
instance
by
vo
ti
ng
of
the
ind
i
vidua
l
trees.
Suppor
t
Vecto
r
Ma
chine
re
volves
a
rou
nd
t
he
noti
on
of
a
m
arg
in
-
ei
t
her
side
of
a
hy
pe
rp
la
ne
that
se
par
at
es
two
cl
asse
s.
Ma
xim
iz
ing
the
m
arg
in
a
nd
wit
h
this
w
ay
creati
ng
t
he
la
rg
est
poss
ible
distance
betwee
n
the
separ
at
in
g
hy
pe
rp
la
ne
an
d
th
e
instances
on
ei
ther
side
of
it
has
bee
n
pro
ve
n
to
re
du
ce
a
n
uppe
r
bo
und
on
the
exp
ect
e
d
ge
ne
rali
zat
ion
er
ror.
RF
wa
s
not
sensiti
ve
to
i
nput
pa
ram
et
e
rs;
thu
s
,
we
just
us
e
d
the
def
a
ult
par
am
et
ers
for
each
cl
assifi
e
r.
T
he
trai
ne
d
cl
assifi
ers
retu
rn
sc
or
e
s
bet
w
een
0
a
n
d
1,
these
sc
or
es
a
r
e
then
trans
form
ed
to
a
bin
a
ry
sta
te
ind
ic
at
in
g
‘
neg
at
ive
’
or
‘
po
sit
ive
’.
For
each
c
om
bin
at
ion
,
the
e
xistence
of
el
e
m
ent
is
consi
der
e
d
posit
ive
(P)
or
ne
gative
(N).
Be
fore
tur
ning
to
pola
rity
,
it
m
ay
be
interest
ing
to
identif
y
wh
et
her
t
he
do
cum
ent
cor
r
es
ponds
t
o
a
s
ub
j
ect
ive
opinio
n
or
a
n
obj
ect
iv
e
fact.
We
w
ould
ha
ve
a
t
w
o
-
ste
p
analy
sis.
O
bjec
ti
vity
an
d
s
ubje
ct
ivit
y as sh
own
in Fi
gure
7.
Figure
7
.
O
bje
ct
ivit
y and
s
ub
j
ect
ivit
y
The
nota
ti
on
of
TP
in
dicat
es
Tru
e
P
os
it
ives:
num
ber
of
exa
m
ples
pr
edict
e
d
posit
ive
that
are
act
ually
po
sit
ive
,
FP
i
ndic
at
es
False
P
os
it
ives:
num
ber
of
exam
ples
pr
e
dicte
d
posit
ive
that
are
ac
tuall
y
neg
at
ive
,
TN
ind
ic
at
es
Tr
ue
Neg
at
ive
s:
num
ber
of
exam
ples
pr
edict
ed
ne
gative
that
ar
e
act
ually
neg
at
ive
and
F
N
in
dicat
es
False
N
e
gative
s: n
um
ber
of e
xam
ples p
re
dic
te
d
ne
gative
th
at
are
act
ually
po
sit
ive
.
The
cl
assifi
cat
ion
m
et
rics
consi
der
e
d
f
or
th
e
senti
m
ent
analy
sis
are
Accur
acy
,
Pr
eci
sio
n,
Re
cal
l
and
F
-
Me
asu
re
a
nd
these
pa
ram
eter
s
are
e
valuat
ed
ba
sed
on
th
e
cal
cul
at
ed
po
sit
ivit
y
and
ne
gativit
y
of
rev
i
ews
by
the
propose
d
hybr
id
a
ppr
oach.
The
pe
rfor
m
ance
eva
luati
on
of
cl
assifi
ers
is
m
a
de
accor
di
ng
to
the
fo
ll
owin
g form
ulas:
Re
port of t
he
tr
ue posi
ti
ves.
It
corres
ponds to:
=
(
+
)
⁄
It
is
thus
the
re
port
bet
wee
n
the
nu
m
ber
of
posit
ive
insta
nc
es
cl
assifi
ed
w
el
l
and
t
he
tota
l
nu
m
ber
of
el
e
m
ents
wh
ic
h
s
hould
be
cl
assifi
e
d
well
.
Re
port
of
the
false
po
sit
ive
on
e
.
He
corres
ponds,
sy
m
m
e
tric
al
ly
in
the
pr
e
vious
de
fini
ti
on
:
=
(
+
)
⁄
The
datum
of
t
he
rates
TP
Ra
te
and
FP
Ra
te
al
lows
t
o
reconstr
uct
the
m
atr
ix
of
co
nfusi
on
f
or
a
giv
e
n
c
la
ss.
Pr
eci
sio
n
is
th
e
repo
rt
betwe
en
the
num
ber
of
t
he
tr
ue
posi
ti
ve
an
d
the
s
um
of
the
tr
ue
posit
ives
a
nd
t
he
false
po
sit
ive
. A
value of
1 ex
pr
e
ss
es
the
fact t
hat
al
l t
he
posit
ive
cl
assifi
ed
e
xa
m
ples w
ere r
ea
ll
y:
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel
Hybri
d
Cl
as
sif
ic
atio
n
A
pproac
h
fo
r
S
e
ntiment
…
(
Ya
ssi
ne
Al A
m
ra
ni)
4561
=
(
+
)
⁄
Re
cal
l
is
the
pe
rcen
ta
ge
of
c
orrect
it
em
s
that
are
sel
ect
e
d.
recall
of
1
m
eans
t
hat
al
l
th
e
posit
ive
e
xa
m
ple
s
wer
e
foun
d.
=
(
+
)
⁄
Accuracy
is
a
com
m
on
m
eas
ur
e for
the
cl
as
sific
at
ion
pe
rfo
rm
ance
and
it
’
s
pro
portion
al
o
f
c
orrectl
y
cl
a
ssifie
d
instances
to
th
e
total
nu
m
ber
of
insta
nces,
wh
er
eas
the
error
rate
use
s
incorrect
ly
classified
rathe
r
than
correct
ly
.
=
(
+
)
(
+
+
+
)
⁄
This
qu
a
ntit
y
al
lows
to
gro
up
in
a
sin
gle
nu
m
ber
the
pe
rfor
m
ances
of
the
cl
assifi
er
(
for
a
giv
e
n
cl
ass)
as
reg
a
rds Rec
al
l
and the
Pr
eci
si
on
:
−
=
(
2
∗
∗
)
(
+
)
⁄
5.
R
ESULT
S
AND DI
SCUS
S
ION
S
To
eval
uate
our
a
ppr
oach,
we
us
e
d
the
"
Am
azon
"
data
set
wh
ic
h
c
onta
ins
1000
inst
ances
di
vid
e
d
into
po
sit
ive
(
500)
a
nd
ne
gative
(
500).
We
di
vid
ed
this
data
into
t
wo
set
s:
a
trai
ning
s
et
and
a
te
st
set
.
I
n
thi
s
arti
cl
e, Cro
s
s
Vali
dation
m
eth
od
with
f
old
va
lue equal
t
o 1
0 has
bee
n use
d for trai
ning a
nd test
ing p
has
es.
We
will
us
e
so
m
e
te
chn
ique
s
that
autom
at
ic
al
ly
extracts
this
data
into
po
sit
ive
or
ne
gative
sentim
ents.
By
us
in
g
th
e
sent
i
m
ent
analy
sis,
the
cu
stom
er
can
know
the
feedbac
k
a
bo
ut
the
pro
duct
be
fore
m
aking
a
purc
hase.
Sentim
ent
analy
sis
is
a
ty
pe
of
nat
ur
al
la
nguag
e
proce
ssing
f
or
trac
ki
ng
t
he
m
oo
d
of
the
public ab
out a
par
ti
cula
r product
.
5
.
1.
Usin
g R
andom
F
orest
Table
1
sho
w
the
res
ult
obt
ai
ned
us
in
g
t
he
Ra
ndom
Fo
r
est
al
gorithm
.
Lo
ok
i
ng
at
t
he
res
ults
of
T
able
1
,
we
no
ti
ce
that
82
0
r
eviews
are
c
orrectl
y
cl
assifi
e
d
am
on
g
1000,
and
18
0
re
views
are
m
isc
la
s
sifie
d.
Figure
8
s
how
the
c
os
t
of
r
a
ndom
f
orest
f
or
cl
ass
p
os
it
ive.
Figure
9
s
how
the
c
os
t
of
r
a
ndom
f
orest
f
or
cl
ass
n
egati
ve
.
Table
1.
C
ro
s
s
Vali
dation R
es
ults f
or Ran
dom
Fo
rest
Po
sitiv
e
Neg
ativ
e
Total
Po
sitiv
e
415
85
500
Neg
ativ
e
95
405
500
Total
510
490
1000
Figure
8
.
Cost
analy
sis o
f
ra
ndom
forest al
gorithm
f
or cla
s
s posit
ive
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2088
-
8708
In
t J
Elec
&
C
om
p
En
g,
V
ol.
8
, N
o.
6
,
Dece
m
ber
2
01
8
:
4554
-
4
567
4562
Figure
9
.
Cost
analy
sis o
f
ra
ndom
f
orest
alg
or
it
hm
f
or cla
s
s n
e
gative
5
.
2
.
Usin
g Su
ppo
r
t Vect
or Mac
hine
Table
2
s
how
t
he
res
ult
obta
ined
us
in
g
S
up
port
Vecto
r
M
achine
al
gorith
m
.
Loo
ki
ng
at
the
re
su
lt
s
of
T
able
2
,
we
no
ti
ce
that
82
4
r
eviews
are
c
orrectl
y
cl
assifi
e
d
am
on
g
1000,
and
17
6
re
views
are
m
isc
la
s
sifie
d.
Figure
1
0
s
ho
w
th
e
c
os
t
of
s
upport
v
ect
or
m
achine
f
or
c
la
ss
p
os
it
ive.
Figure
1
1
s
ho
w
th
e
c
os
t
of
s
upport
v
ect
or
m
achine for
class
n
e
gat
ive.
Table
2
.
Cr
oss
Vali
dation R
es
ults f
or S
uppor
t Vecto
r
Ma
chi
ne
Po
sitiv
e
Neg
ativ
e
Total
Po
sitiv
e
409
91
500
Neg
ativ
e
85
415
500
Total
494
506
1000
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Elec
&
C
om
p
En
g
IS
S
N: 20
88
-
8708
A Novel
Hybri
d
Cl
as
sif
ic
atio
n
A
pproac
h
fo
r
S
e
ntiment
…
(
Ya
ssi
ne
Al A
m
ra
ni)
4563
Figure
10
.
C
ost
an
al
ysi
s of s
uppo
rt v
ect
or
m
achine al
gorith
m
f
or
class
pos
it
ive
Figure
11
.
C
ost
an
al
ysi
s of s
uppo
rt v
ect
or
m
achine al
gorith
m
f
or
class
negat
ive
5
.
3
.
Usin
g R
andom
F
orest
Support
V
ec
t
or Mac
hine
Table
3
s
how
the
res
ult
obt
ai
ned
us
i
ng
our
ap
proac
h
R
andom
Fo
re
st
Sup
port
Vect
or
Ma
chi
ne
al
gorithm
(
RFSVM
)
.
Lo
o
king
at
the
res
ults
of
T
a
ble
3
,
w
e
no
ti
ce
that
847
rev
ie
ws
are
correct
ly
cl
assifi
ed
a
m
on
g
1000,
a
nd
153
re
view
s
are
m
isc
la
ssi
fied.
Fi
gure
1
2
sh
ow
the
Cost
of
Ra
ndom
Fo
rest
S
upport
Vecto
r
Ma
chine
f
or
cl
ass
Po
sit
ive.
F
igure
1
3
s
how
the
Cost
of
Ra
ndom
Fo
rest
Suppor
t
Vecto
r
Ma
chine
for
cl
ass
Neg
at
ive
.
Table
3
.
C
ro
s
s
Vali
dation R
es
ults f
or RFS
V
M
Po
sitiv
e
Neg
ativ
e
Total
Po
sitiv
e
422
78
500
Neg
ativ
e
75
425
500
Total
497
503
1000
Evaluation Warning : The document was created with Spire.PDF for Python.