TELKOM
NIKA
, Vol. 11, No. 7, July 201
3, pp. 3534 ~ 3540
e-ISSN: 2087
-278X
3534
Re
cei
v
ed
Jan
uary 28, 201
3
;
Revi
sed Ma
rch 1
3
, 2013;
Acce
pted Ma
rch 3
0
, 2013
Effect of User’s Judging Powe
r on the
Recommendation Performance
Li-Yu Mao, Yuan Guan, M
i
ng-Sheng S
h
ang*, Shi-M
i
n Cai
W
eb Scienc
es Center, Scho
ol
of Computer S
c
ienc
e an
d En
gin
eeri
ng, Un
iv
ersit
y
of El
ectronic Sci
ence
a
nd
T
e
chnolog
y
of Chin
a, Che
n
g
d
u
611
73
1, P.R. Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: shang.mi
ngs
hen
g@gm
ail.c
o
m
A
b
st
r
a
ct
In most B
2
C E
-
commerce websites, rec
o
mm
en
der system
s
make recomm
endations
f
o
r eac
h
indiv
i
d
ual
user
base
d
o
n
h
i
s
/
her histor
ical
rating
be
havi
o
rs. Previous
lit
eratures foc
u
s
on th
e ov
era
ll
perfor
m
ance of
recommender
system
, while
the perfor
m
anc
e
of indiv
i
dual level rec
e
iv
es little attention. In
this pa
per, w
e
discov
e
r that reco
mme
ndati
o
n alg
o
rith
ms
p
e
rform
better o
n
users w
ho h
a
ve stron
g
ju
d
g
in
g
pow
er, an
d vic
e
versa. W
e
te
st our conc
lusi
on o
n
thre
e
b
e
n
ch
mark data sets,
na
mely Movie
Lens,
N
e
tflix,
and
A
m
a
z
o
n
,
w
h
ich further
provi
de
evid
en
ce of th
e v
a
li
d
i
ty of o
u
r fin
d
i
ng. Mor
eover,
our
find
in
g
may
provi
de s
o
me
gui
danc
e for
d
e
sig
n
in
g rec
o
mme
n
d
a
tion
al
gorith
m
s
more
efficiently
by
concer
nin
g
us
ers'
different ju
dg
in
g pow
er.
Ke
y
w
ords
:
Re
commen
der sy
stem, user'
s
ju
dgi
ng p
o
w
e
r, colla
bor
ative filt
erin
g, RMSE
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introducti
on
The
rapi
d
de
velopment
of
Intern
et tech
nology
ma
ke
s u
s
en
co
unt
er l
o
ts
of info
rmation,
e.g., tens of thou
san
d
s of
movies in
Net
f
lix, m
illions o
f
books in Am
azo
n
, over o
n
e
billion of
we
b
page
s collect
ed by delicio
us.com and
so on. How to
find the interesting p
a
rt a
m
ong them i
s
a
big chall
eng
e
.
The
t
r
aditio
nal sea
r
ch e
ngine
can
on
ly pre
s
e
n
t all
users
with
the
sam
e
re
sults,
while
it cann
ot provid
e p
e
r
so
nali
z
ed
se
rvice
s
co
n
c
e
r
ning
different
users’ inte
re
sts
and
ho
bbi
es
[1]. With su
ch add
re
ssed
issue, the p
e
r
so
nali
z
ed
re
comm
end
atio
n se
rvice spri
ngs
up, an
d
has
been inve
stig
ated extensiv
ely [2] [3]
[4]. In
most B2C e-comm
erce web
s
ites,
recommen
d
e
r
system
s re
co
mmend o
b
je
cts for use
r
s b
a
se
d on thei
r past online b
ehaviors (e.g. click, bro
w
se
,
purcha
s
e
)
an
d use
r
s a
r
e n
o
longe
r pa
ss
ive browse
rs but active participant
s.
A variety of person
a
lize
d
re
comm
en
dation al
gori
t
hms have
been
pro
p
o
s
ed by
resea
r
chers,
inclu
d
ing
coll
aborative filtering
meth
od
s [5] [6],
con
t
ent-ba
s
ed
m
e
thod
s [7]
a
nd
hybrid o
n
e
s
[8]. Howeve
r, previou
s
work ma
inly f
o
cu
se
d on t
he overall p
e
rform
a
n
c
e
s
o
f
recomme
ndat
ion alg
o
rithm
s
, while p
a
id
little attention on t
he
re
commen
dation
perfo
rma
n
ce
of
individual lev
e
l. But in the real
ca
se
s,
there
a
r
e v
a
riou
s
kind
s
of use
r
s, a
n
d
they may rate
obje
c
ts in different ways. F
o
r exampl
e, the user may
be som
eon
e who d
o
e
s
not
taken seri
ou
sly
about voting,
or he/
she
ha
s no
expe
rien
ce in th
e
rela
ted field an
d
gives
some
irrational
ratin
g
s
.
What’
s
worse,
some mali
ciou
s spam
m
e
rs give
bia
s
ed ratin
g
s i
n
tentionally. We su
ppo
se th
at
use
r
s have
di
fferent jud
g
in
g po
we
r, whi
c
h
may h
a
ve
some
certai
n
impact
s
o
n
th
e pe
rform
a
n
c
es
of reco
mmen
dation alg
o
rit
h
ms.
There
have b
een some ra
nkin
g
algo
rith
ms
wh
ich ca
n be u
s
e
d
to disting
u
ish
u
s
ers by
their ju
dgin
g
power [9] [10] [11]. For example, in
Ref
s
[9] [10
], an iterativ
e refin
e
ment
(IR)
algorith
m
is p
r
opo
se
d. In [11], the iterative re
fineme
n
t
algorithm is
revise
d by De Kerchove a
n
d
Van Do
oren,
whi
c
h a
s
sign
trust to ea
ch i
ndividual
rati
ng. In this p
a
per, we p
r
op
ose
an alg
o
rit
h
m
based o
n
YZ
LM (Yu
-
Zh
an
g-La
ureti-Mo
ret, see i
n
Re
f
[9]) to me
asure
users' j
u
d
g
ing p
o
wer. T
he
cla
ssi
c user-based coll
ab
orative
filteri
ng m
e
thod
(CF) is u
s
ed
to te
st the
recomme
ndat
ion
perfo
rman
ce
on differe
nt use
r
s
with
different jud
g
i
ng po
we
r. We first divide all u
s
ers into
different g
r
ou
ps by thei
r ju
dging
po
wer.
Then
we g
e
t the avera
ge
intro-gro
up
re
comm
end
atio
n
perfo
rman
ce
of differe
nt g
r
oup
s. Th
rou
g
h
exten
s
ive
e
x
perime
n
ts o
n
three ben
chmark data
sets,
w
e
find
th
a
t
C
F
pe
r
f
or
ms
b
e
tte
r
on
us
er
s
w
h
o
ha
ve
s
t
r
o
ng
er
ju
dg
in
g
po
w
e
r
,
vice
ve
rs
a
.
In
o
t
h
e
r
words, it sho
w
s th
at the a
c
cura
cy pe
rfo
r
man
c
e
on
e
a
ch
user i
s
p
o
sitively co
rre
l
ated with
his/
her
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Effect of user
’
s
judgi
ng po
wer on the recom
m
endation perform
an
ce
(Li-Yu Ma
o)
3535
judgin
g
powe
r
. Moreove
r
, our findin
g
m
a
y provide some guid
a
n
c
e for desi
gni
ng more efficient
recomme
ndat
ion algo
rithm
s
co
ncerni
ng
each user'
s
ju
dging p
o
we
r.
This pa
per ta
ckl
es the
imp
o
rtant i
s
sue
o
f
how
to
mea
s
ure th
e u
s
e
r
'
s
ju
dgin
g
p
o
w
er,
an
d
find that user'
s
jud
g
ing
po
wer i
nde
ed h
a
ve a po
sitive co
rrelation i
m
pact o
n
the
re
comme
nd
ation
perfo
rman
ce.
The p
ape
r
is o
r
gani
ze
d
as follo
ws. Section
1 i
s
the int
r
od
uction
part.
The
recomme
ndat
ion alg
o
rithm
and m
odified
YZLM alg
o
rit
h
m are introd
uce
d
in
se
cti
on 2. Se
ction
3
is abo
ut the experime
n
tal d
a
ta sets a
nd the re
sult
s
we
get. The last part is ou
r co
nclu
sio
n
s.
2. Algorithms and Metric
s
2.1.
Us
er-
B
a
sed CF
Algo
rithm
The ba
sic id
ea of CF algorithm
can
be di
vided i
n
to two step
s: (1) Calcul
ate the
simila
rities b
e
twee
n the
t
a
rget
u
s
er a
nd hi
s/he
r
ne
ighbo
rs throu
gh thei
r
histo
r
y be
haviors;
(2
)
Predi
ct the target user'
s
p
r
eferen
ce fo
r an uno
bserve
d obje
c
t.
(1) Th
e deg
ree of similari
ty between u
s
er
u a
nd u
s
er v is measured by form
ula (1
)
through
c
o
s
i
ne s
i
milarity. Here,
r
u,i
is th
e e
x
is
tin
g
r
a
ting
o
f
us
er
u to o
b
j
ec
t i, an
d I
u
denote
s
th
e
set of obje
c
ts rated by use
r
u.
u
v
u
v
u
,
i
v
,
i
i
︵
I
I
︶
2
2
u
,
i
v
,
i
i
I
i
I
r
*
r
s
i
m
︵
u
,
v
︶
r
*
r
(1)
(2) T
he predi
cted value of
the target u
s
e
r
u to the obje
c
t i is cal
c
ulat
ed by
u
u
v
,
i
v
.
v
N
u
,
i
u
.
v
N
s
i
m
︵
u
,
v
︶
*
︵
r
r
︶
p
r
s
i
m
︵
u
,
v
︶
(2)
whe
r
e t
he
col
l
ection
N
u
i
s
the
set of
u
s
e
r
s who
also rated the
obj
e
c
t i.
r
v.
is the
averag
e
ratin
g
given by use
r
v.
2.2.
Modified
YZML Algor
ithm
Most we
bsite
s
su
ch a
s
Amazo
n
, MovieLen
s
and
Netflix usually use the a
r
ith
m
etical
averag
e of the obje
c
t’s rati
ngs a
s
the estimation of
its quality. However, it does
not con
s
id
er the
differen
c
e
s
o
f
use
r
s’ j
udgi
ng po
we
r. Y
Z
LM al
g
o
rith
m makes a
distin
ction b
e
t
ween
users
with
different p
r
ofil
es by th
eir ju
dging
po
wer,
whi
c
h i
s
p
r
o
portion
al to u
s
ers’
weig
hts.
Users’ ju
dgi
ng
power is the
n
used by the
weig
hted arit
hmetic ave
r
a
ge to es
timat
e
for the objec
t’s
quality. In
this
way, we can
get a more a
c
curate estim
a
tion for the ob
ject’s q
uality.
Suppo
se
N u
s
ers a
nd M o
b
ject
s in a rating sy
ste
m
. Each u
s
e
r
ha
s his/he
r own j
udgin
g
power (d
enot
ed by
W
u
fo
r
rater
u, l
a
rg
e
r
W
u
corre
s
p
ond
s to
stron
ger judgi
ng
p
o
we
r) an
d e
a
c
h
obje
c
t has
a
n
intrin
sic q
u
a
lity (indicate
d by Q
j
for o
b
ject j).
We
assume th
at both the jud
g
i
n
g
power a
nd int
r
insi
c q
uality are late
nt.
σ
2
,
u
repre
s
e
n
ts the
deviation of the ratin
g
vect
or of u
s
er
u
from the obje
c
t's qu
ality vector, and it ha
s an inverse
correl
ation wi
th W
u
. The Q
j
and
σ
2
,
u
will be
estimated by
q
j
and V
u
. In YZLM alg
o
r
ithm, the obj
ect’s
quality i
s
e
s
timated
by the wei
g
h
t
ed
arithmeti
c
av
erag
e, wh
ere
the
wei
ghts
is propo
rtion
a
l to user
s’ j
udgin
g
po
we
r and th
e u
s
ers’
judgin
g
po
we
r are up
date
d
by the e
s
timated obj
ec
t
s
’ qu
ality. By iterative refi
nement, we
can
obtain the
q
j
and
V
u
a
s
cl
o
s
e a
s
p
o
ssibl
e
to the hi
dd
en value
s
Q
j
and
σ
2
,
u
after
converg
e
n
c
e o
f
the algorith
m
.
The ori
g
inal i
m
pleme
n
tatio
n
of YZLM algorithm
con
s
i
ders only the
case wh
en a
ll users
have rate
d all
object
s
, whil
e it cann
ot be
gene
rali
zed t
o
handl
e sparse d
a
ta. In order to p
r
o
c
e
s
s
s
p
ar
se
da
ta
,
w
e
u
s
e
A, an
N
×
M ad
jac
e
nc
y ma
tr
ix, to
r
e
c
o
r
d
the spare data. If rater
u rat
e
obje
c
t j, A
uj
=1, otherwi
se A
uj
=
0
[12].
Each
user u i
s
a
s
sign
ed
with a
weight v
a
lue
w
u
, which is initially set as 1/N.
r
u,
j
is the
rating
user
u
rate
s to o
b
j
e
ct j. Th
e qu
ality of
obje
c
t j is
estimate
d by the
wei
ghted a
r
ithm
etic
averag
e.
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3534 – 354
0
3536
N
j
u
j
u
u
,
j
u
1
q
A
w
r
(3)
The ratin
g
varian
ce of user u
is com
pute
d
as follo
ws
u
,
i
M
2
2
j
u
u
u
,
j
u
,
j
M
j
1
i
1
1
V
A
︵
r
q
︶
A
(4)
It should b
e
noted that th
e
σ
2
,
u
som
e
tim
e
s may e
qua
l to 0. Thus,
we
con
s
trai
n
th
e
value of
σ
2
,
u
to be not le
ss than a
ce
rtain small va
lue
ε
>0 to p
r
event u
s
er
weig
hts from
divergin
g (In
our
simulatio
n
s, we use
ε
=10
-8
). T
he u
pdated n
o
rm
alize
d
wei
ght
of use
r
u is t
hen
given by
-
u
u
N
-
v
v
1
V
w
V
(5)
whe
r
e
β≥
0. It is very obvious that
β
=0
corre
s
po
nd
s to the simpl
e
arithmeti
c
averag
e.
The highe
r
β
will
bri
ng the greater penalization
to
users
with larger deviation V
u
. Yu et
al. [9]
noted th
at the
ca
se
β
=1/2 provides
better nu
merical stability of
the
algorit
h
m
as well as
transl
a
tional
and
scale inv
a
rian
ce,
while
in Ref [13], t
he case
β
=1 i
s
the o
p
timal
from the p
o
int
of
view of
math
ematical
stati
s
tics. Herein,
we
use
β
=1
b
e
ca
use it yiel
ds
su
peri
o
r p
e
rform
a
n
c
e,
and
t
he ca
se
β
=1/2 doe
s not alter the funda
mental ch
ara
c
ter of final re
sult.
The algo
rith
m is initialize
d
by setting the use
r
wei
ghts a
s
w
u
=1/N for all users
,
then
iterates
rep
e
a
tedly over the equ
ation
s
(3, 4, 5)
unti
l
the chan
ge
of the estima
ted quality vector
betwe
en two
adja
c
ent itera
t
ion step
s is l
e
ss than a ce
rtain thre
sh
ol
d value
∆
.
k
2
j
j
j
︷
k
|
q
0
︸
k
1
|
q
q
'
|
︵
q
q
︶
|
︷
k
|
q
0
︸
|
(6)
Note th
at the
algo
rithm m
a
y fail to con
v
erge if th
e v
a
lue of th
re
shold
∆
i
s
set
to be too
small.
Conve
r
sely, too l
a
rge value
of thre
shol
d
may
disrupt the it
erative p
r
o
c
e
ss [9]. Th
erefore
it's better to take a fe
w trials to ch
oo
se
an approp
ria
t
e value, and
the value is set as
∆
=10
-6
in
our s
i
mulations
[14].
2.3.
Perform
a
nce Me
trics
The
accu
ra
cy metric is often u
s
e
d
to me
asure the
pe
rformanc
es of
different
recomme
ndat
ion algo
rithm
s
[15]. The
mean a
b
solu
te erro
r (MA
E
) is a wi
del
y used a
c
curacy
metric th
at compute
s
the
mean
ab
solut
e
deviati
on
o
f
two sequ
en
ce
s. The
MAE of use
r
lev
e
l,
MAE
u
, is
c
a
lculated as
follows
:
u
u
,
i
u
,
i
i
T
u
u
p
r
M
A
E
T
(7)
whe
r
e p
u,i
i
s
the predi
cted
rating g
ene
ra
t
ed by the alg
o
rithm of
CF. r
u,i
is the
act
ual ratin
g
u
s
e
r
u
gives to obje
c
t i in the probe set, and |T
u
| is the numb
e
r of rating
s o
f
user u in the
probe
set.
The un
evenn
ess of the weights a
s
sign
ed to
individ
ual user
can
be mea
s
u
r
e
d
by the
inverse pa
rticipation ratio
(IPR).
Given th
e use
r
no
rmal
ized
weig
hts
w
u
, IPR can b
e
comp
uted a
s
2
1
u
u
U
I
P
R
︵
w
︶
(8)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Effect of user
’
s
judgi
ng po
wer on the recom
m
endation perform
an
ce
(Li-Yu Ma
o)
3537
The IPR i
s
reci
procal to
anoth
e
r we
ll-kn
own m
e
asu
r
e, th
e
Herfin
dahl
-Hi
r
schm
an
con
c
e
n
tration
index (HHI). The IP
R mea
s
ures the effe
ctive numbe
r
of
users with
respe
c
t to their
weig
hts w
u
.
Whe
n
all
wei
ghts
are e
q
u
a
l, w
u
=1/N, th
en IPR=N. B
y
cont
ra
st, when
all
weig
h
t
s but
one are ze
ro,
IPR=1.
3.
Experime
ntal Re
sults
3.1.
Da
tas
e
t
s
Thre
e ben
ch
mark data sets are u
s
e
d
to test the algori
t
hms’ pe
rform
ances:
(1) MovieLens
(http://www.movielens
.org) is
a movie recommendation webs
ite, which
use
s
u
s
e
r
s'
rating
s to gene
rate p
e
r
so
nali
z
ed
recom
m
en
dati
ons. Th
e d
a
ta we u
s
e
d
is
downloaded f
r
om http://www
.grouplens
.
org/node/73.
(2)
Netflix
(http://www.ne
tix
.
com) pr
ov
ides the
wo
rld'
s lar
g
e
s
t online v
i
de
o renta
l
servi
c
e, offeri
ng mo
re tha
n
6.7 million
subscri
b
e
r
s
acce
ss to
85,00
0 DVD and
a
gro
w
ing li
bra
r
y
of over 4,000
full-length m
o
vies an
d television e
p
iso
des that a
r
e
available for i
n
stant watchi
ng
on their PCs. The data we use
d
is a ra
n
dom sa
mp
le that con
s
i
s
ts o
f
3000 use
r
s who h
a
ve rat
ed
at least 20 m
o
vies an
d 27
79 movies h
a
v
ing been rated by at least
one user.
(3) Amazon (http://www.amazon.
com)
is a multinati
onal
e-comm
erce company. Th
e
origin
al
data
were colle
cte
d
from 28 Jul
y
2005
to
27
Septembe
r 2
005, an
d the
data we u
s
e
d
is a
rand
om samp
le.
The ba
si
c sta
t
istics
of thre
e ben
chm
a
rk
data sets a
r
e
sho
w
n in
Tab
l
e 1, in whi
c
h
we
can
find that they
have differe
nt sizes an
d
different
spa
r
sity. In orde
r to co
mprehe
nsively test t
h
e
recomme
ndat
ion perfo
rma
n
ce, the data
are ran
doml
y
divided into two parts: th
e 80% trainin
g
set (E) an
d the 20% pro
be set (T
). The inform
ation of trainin
g
set is trea
ted as kn
own
informatio
n, while n
o
information of pro
be set is all
o
wed to u
s
e fo
r pre
d
ictio
n
.
Table 1. Basi
c statisti
cs
of
the tested dat
a sets
Data set
User
s
O
b
jects
Ratings
Spar
sity
MovieLens 943
1683
100000
6.30×10
-
2
Netflix 3000
2779
197248
2.37×10
-
2
Amazon 3604
4000
134679
9.34×10
-
3
Furthe
rmo
r
e,
the distributi
ons of ratin
g
s
are
sh
o
w
i
n
Figure 1. It is intere
sting
that th
e
data of MovieLen
s and Netflix share the simila
r pa
ttern, and differ from the data of Amazon. T
he
main re
ason
may be that the MovieLe
ns an
d Netfli
x
only includ
e the media
obje
c
ts, and
it
make
s the
user
ea
sily com
pare
the
qu
alities of
differe
nt obje
c
ts.
Bu
t the Ama
z
on’
s d
a
ta i
s
highl
y
spa
r
se and A
m
azo
n
’s u
s
e
r
s only buy/rat
e what they a
c
tually like.
Figure 1.
Di
st
ribution
s
of ra
tings in thre
e data set
s
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3534 – 354
0
3538
3.2.
Re
sults &
Disc
ussio
n
s
To investig
ate the effect o
f
users’ ju
dgi
ng po
we
r on
recomme
ndat
ion pe
rform
a
nce,
we
need to co
mpute their
judgin
g
power by modi
fied YZML algorithm. Fig
u
re 2 sh
ows the
histog
ram
s
o
f
σ
u
's di
stri
b
u
tions for th
e thre
e d
a
ta
set, in
whi
c
h
three
subpl
ots
rep
r
e
s
ent
the
results of M
o
vieLen
s, Netf
lix and Amazon, re
spe
c
ti
vely. In three
datasets, the
cente
r
pe
aks of
histog
ram
s
a
r
e aro
und 1 o
r
0.75, whi
c
h
sug
g
e
s
ts t
hat
most users’ ratings diffe
r from the obje
c
t
s
’
qualities.
0.5
1
.
0
1.5
2
.0
0
20
40
60
80
100
nu
mber o
f
u
s
ers
e
s
ti
ma
ted
M
o
vieLen
s
0.5
1
.0
1.5
2
.
0
0
50
100
150
200
250
nu
mber o
f
u
s
ers
es
tim
a
ted
N
e
t
f
lix
0
.
51
.
0
1
.
52
.
0
2
.
5
0
50
100
150
200
nu
mber o
f
u
s
ers
e
s
ti
ma
ted
A
m
az
on
Figure 2.
Hi
st
ogra
m
s of
σ
u
'
s
dist
ribution
s
by modified YZLM with
β
=1
The val
ue
of IPR
clo
s
ely
relate
s to
th
e hi
stog
ram
s
of
σ
u
's
d
i
s
t
r
i
b
u
t
io
ns
acc
o
r
d
in
g to
equatio
n (8).
The IPR valu
es
dep
end
ent
on
β
a
r
e
su
mmari
zed
in
Table
2 fo
r al
l the th
ree
da
ta
s
e
ts
. The
c
a
se
β
=0
ma
ke
s the hom
oge
neou
s
weig
ht of use
r
s, whi
c
h le
ad
s to th
e value
s
of IPR
euqal to the total numbe
r
of use
r
s. As
β
increa
se
s, IPR value gradua
lly de
cre
a
se
s, indi
cati
ng
that YZLM algorithm can
distingui
sh
betwe
en
different u
s
ers. User
s of poor judgi
ng p
o
we
r
grad
ually lose the right to spea
k. We
can find
that
the IPR value of
sparse
r dataset
decli
nes
faster,
espe
ci
ally in the
A
m
azo
n
. IPR
value d
r
op
s t
o
13
when
β
=1. Fi
gure 2
sho
w
s that t
h
e
numbe
r of u
s
ers in the first
bin of Ama
z
o
n
(i.e., with
σ
i
clo
s
e to
ze
ro
) is
13. Thi
s
v
a
lue i
s
eq
ual
to
the value of I
P
R. The
s
e
“i
deal u
s
e
r
s”
with small
esti
mated
σ
u
≈ε
have ver
y
large w
u
≈
1/
ε
(a s
m
all
ε
=10
-8
is ch
osen as a lo
wer bound to avo
i
d the diverg
e
n
ce of user weights). This
doe
s not mea
n
that the
effective numb
e
r
of users i
s
1
3
an
d
ra
ting
s of the
othe
r users
are n
egle
c
ted. Oth
e
r
use
r
s
cou
n
t for all obje
c
ts that have not been rat
ed
by the “ideal
use
r
s”. Besid
e
s, these “id
eal
use
r
s” corre
s
pond to users with a few ra
tings (n
ear 8 f
o
r
β
=1). In an
extreme ca
se, if a user only
rates a
n
obje
c
t and this o
b
j
ect is only rat
ed by him, his estim
a
ted
σ
u
=
ε
.
Table 2. IPRs for three dat
a sets
with different
β
Dataset
β
=0
β
=0.5
β
=1
β
=1.5
β
=2
β
=2.5
β
=3
MovieLens 943
895
762
557
209
2
3
Netflix 3000
2833
2276
373
5
5
7
Amazon 3604
1242
13
127
140
137
136
With the in
cre
a
se
of
β
, “ide
al users”
gra
d
ually app
eare
d
in Movielens and Netflix.
This i
s
becau
se YZL
M
algorithm i
s
a proc
ess
of iterative re
finement. Wh
en
β
is large
enoug
h, these
use
r
s
who in
the first iteration step with
rathe
r
small
estimated
σ
u
are given la
rge weig
ht in the
se
con
d
iterati
on step. The
n
, thes
e users’ rating
s hav
e the right to
spe
a
k to the estimated q
u
a
lity
values an
d f
u
rthe
r lo
we
r t
heir estim
a
te
d
σ
u
. By re
p
eating th
ese
iteration
s
,
σ
u
became
sm
al
ler
and sm
alle
r. Finally, there
may appea
r some “id
eal u
s
ers” with e
s
ti
mated
σ
u
≈
ε
.
We
equally
divide all
users into te
n
gro
u
p
s
a
ccordin
g to th
eir ju
dgin
g
p
o
we
r by
desce
nding
o
r
de
r. Each g
r
oup con
s
ist
s
of about 10%
of the total numbe
r of use
r
s. We obtain
all
use
r
s’ MAE
u
and
then co
unt
avera
ge the
MAE
u
of ten grou
ps.
Figure 3 is a
comp
ari
s
on
of
averag
e MAE
u
of ea
ch
gro
up a
nd th
e a
v
erage
MAE
u
of all
users.
The h
o
ri
zo
ntal axis indi
ca
tes
the averag
e judgin
g
power (
denote
d
by grou
p'
s avera
ge
σ
u
, smaller
σ
u
corre
s
po
nds to strong
er
judgin
g
p
o
we
r)
of the
user group
s, an
d
t
he ve
rtical
axis d
enote
s
the ave
r
a
g
e
MAE
u
of u
s
ers.
The
circle
lin
e an
d
solid
li
ne
rep
r
e
s
ent
the intra-g
r
ou
p ave
r
age
M
A
E
u
and th
e
averag
e MAE
u
of
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
Effect of user
’
s
judgi
ng po
wer on the recom
m
endation perform
an
ce
(Li-Yu Ma
o)
3539
all use
r
s, re
spe
c
tively. The su
bplot
s sho
w
the re
sults of Movi
eLen
s, Netfli
x
and Amazon,
respe
c
tively. In Fig
u
re
3, from left to
righ
t, we
can
find
that the
aver
age j
udgi
ng
p
o
we
r of
grou
ps
is g
e
tting
wo
rse
and
the
av
erag
e MAE
u
of group
s
em
erge
in
a
s
cen
d
ing t
r
en
d. O
v
erall, the
r
e
i
s
a
positively co
rrelated
relati
onship bet
ween CF’s
M
AE and gro
up’s ju
dgin
g
powe
r
, i.e., the
recomme
ndat
ion pe
rform
a
nce i
s
relatively good i
n
the group
wit
h
better ju
dgi
ng po
we
r , a
nd
these users with
po
or
ju
dgi
ng
power have a g
r
e
a
t impact o
n
the perfo
rm
ance of the
CF
algorith
m
.
Figure 3.
Average MAE
s
of group
s an
d a
v
erage MAE
s
of all users for thre
e data
s
ets
Re
comm
end
er
system
ca
n’t meet all
u
s
ers’
dem
and
s if it treats a
ll users e
qual
ly. The
grou
p’s MAE
u
is smalle
r if it consist
s
of users with
b
e
tter judgin
g
power, whi
c
h
means that the
predi
cted
rati
ngs g
ene
rate
d by reco
mm
endatio
n algo
rithm are
clo
s
er to the actu
al rating sco
r
e
.
But the MAE
u
of group
s
with lo
wer j
u
dging
power
are
re
latively highe
r, whi
c
h mean
s that
the
predi
cted
rati
ngs g
ene
rate
d by the CF
are q
u
ite
different fro
m
the actual
scores. The
s
e u
s
ers
with hi
gh
MAE
u
are
suspe
c
t to
rai
s
e th
e
average
MA
E
u
of all
users
(MAE of
sy
stem l
e
vel). S
i
nce
CF
can
not recom
m
en
d a
c
curately fo
r use
r
s with
poor jud
g
ing
power,
we
can
utilize
ot
her
recomme
ndat
ion alg
o
rithm
s
whi
c
h a
r
e
more
in lin
e
with the
u
s
ers' b
ehavio
rs
to make the
s
e
use
r
s’ MAE
u
decrea
s
e. By
this way, the averag
e MA
E
u
of all users will
be improved. That m
a
y
be our
sub
s
e
quent re
se
arch conte
n
t.
4. Conclusi
on
In this pap
er,
we firstly pro
pose a natu
r
al ex
tensio
n
of the YZLM
algorith
m
to g
e
t use
r
s’
judgin
g
p
o
we
r. The
n
we
st
udy the Int
r
a-grou
p p
e
rfo
r
mances of
CF alg
o
rithm
i
n
ten
group
s
with
different judgi
ng po
wer. Th
roug
h expe
ri
ments o
n
three ben
chm
a
rk data
s
ets,
it sho
w
s that there
is a
po
sitively correl
ated
relation
ship
b
e
twee
n u
s
e
r
s’ judgin
g
p
o
wer a
nd th
e re
comm
end
atio
n
perfo
rman
ce.
That's to say, users wit
h
stron
g
jud
g
ing po
we
r are mo
re likely to get b
e
tter
recomme
ndat
ion pe
rform
a
nce,
while th
e use
r
s wh
o
have poo
r ju
dging p
o
wer
can o
n
ly get bad
recomme
ndat
ion re
sults. Besid
e
s, the CF algorithm
p
r
edi
cts the ta
rget user'
s
preferen
ce b
a
sed
on p
r
efe
r
en
ces
of hi
s/her
neigh
bor
s. Users with better
ju
dgin
g
p
o
w
er
a
c
cords with
mai
n
st
re
am
prefe
r
en
ce
s,
so th
e recom
m
endatio
n p
e
r
forma
n
ce i
s
relatively bett
e
r. On
the
co
ntrary, for u
s
ers
with p
oor jud
g
ing
po
wer,
their p
r
efe
r
e
n
ce
s
are
mo
re p
e
rso
nali
z
ed a
nd
hard
to be
ha
ndl
ed.
More
over, sin
c
e CF
cann
ot
sati
sf
y preferences for u
s
e
r
s with
po
or j
udgin
g
p
o
we
r, we
ca
n ta
ke
other alg
o
rith
m to cover CF's sho
r
tage.
This may be
our furth
e
r
studie
s
.
Referen
ces
[1]
Brin S
an
d P
a
ge
L. T
he an
atom
y of
a l
a
rg
e
-
scale
h
y
p
e
rte
x
tu
al
w
e
b
s
e
a
r
ch e
ngi
ne.
Com
p
ut. Netw.
ISDN Syst
. 1998; 30: 10
7-11
7.
[2]
Resnick P and
Varian
H R. Recommender s
ystems.
Commu
n
. ACM
. 1997; 40(3): 56-
58.
[3]
Abeer Mo
ham
ed El-kor
an
y
and Sa
lma M
o
khtar Khat
ab
. Ontolog
y
b
a
s
ed Soci
al Re
commen
der
Sy
s
t
e
m
.
IAES International Journal of Arti
ficial Intelligence (
I
J-AI).
2012; 1(3): 127-1
38.
[4]
Muhamm
ad W
a
seem C
hug
ht
ai, Ali Bin Se
la
mat and
Imran
Ghani. Goal-
b
a
s
ed h
y
br
id filter
ing for user-
to-user Perso
n
a
lize
d
Rec
o
m
m
end
ation.
Int
e
rnati
ona
l Jour
nal of El
ectr
ica
l
and
Co
mp
ute
r
Engin
eer
in
g
(IJECE)
. 2013;
3(3).
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN:
2087
-27
8
X
TELKOM
NIKA
Vol. 11, No
. 7, July 2013
: 3534 – 354
0
3540
[5
]
R
e
sn
i
c
k P, Ia
co
vo
u N
,
Su
cha
k
M, Be
rg
stro
m P an
d R
i
ed
l
J. Gro
u
p
Len
s: An
o
p
e
n
arch
i
t
e
c
tu
re fo
r
colla
bor
ative fil
t
ering of net
ne
w
s
.
In ACM CS
CW
. 1994; 17
5
-
186.
[6]
Lin
den G, Smit
h B, and Y
o
rk
J. Amazon.co
m Re
comme
n
datio
ns Item-to-Item Colla
bor
ative F
ilteri
ng.
IEEE Internet Com
p
uting
. 2
0
03; 7(1): 76-8
0
.
[7]
Pazzan
i
MJ and Bil
l
sus D.
Content-bas
ed recommendat
ion system
s
. T
he Adaptive
W
eb. Berlin
.
Heid
el
berg. 2
0
07; 432
1: 325-
341.
[8]
R Burke.
H
y
bri
d
recomm
en
de
r s
y
stems: S
u
rve
y
an
d e
x
per
i
m
ents.
User M
ode
lin
g a
nd
U
s
er-Ada
pte
d
Interaction
. 2
0
02; 12(4): 3
31-
370.
[9]
Yu YK, Z
h
a
n
g
YC,
Laur
eti
P, an
d Mor
e
t L. Dec
odi
ng
informati
on
from n
o
is
y, re
d
und
ant, a
n
d
intenti
o
n
a
ll
y d
i
st
orted source
s.
Physica A
. 2006; 37
1: 732-
744.
[10]
Laur
eti P, Moret L, Z
hang Y
C
, and Yu YK.
Information fil
t
ering vi
a itera
t
ive refinem
ent
.
EPL
. 2006;
75(6): 10
06-
10
12.
[11]
De Kerch
o
ve
C, Van D
oor
en
P. Iterative f
ilterin
g for a d
y
n
a
mical re
putati
on s
y
st
em.
arXiv 071
1. 396
4
,
200
7.
[12]
De Kerch
o
ve C
and Van D
oor
en P.
Reput
ati
on s
y
stems an
d non
ne
gativit
y.
LNCIS
. 2009;
389: 3-16.
[13]
Hoel PG. Introductio
n
to Mathemat
ic
al Stati
s
tics. W
ile
y
,
N
e
w
Y
o
rk, 198
4.
[14]
Medo C M, W
a
keli
ng J, Mir
y
l
enka RK, et al.
Model of revi
e
w
ers
’
b
ehav
ior
in peer rev
i
ew
s
. Europea
n
Commun
i
t
y
.R
e
port numb
e
r: 2133
60. 20
10.
[15]
Vozal
i
s
E and
Margaritis
KG.
Analys
is of r
e
c
o
mmen
der sys
tem
alg
o
rith
ms
. In Proce
edi
n
g
s of th
e 6t
h
Hell
en
ic Eur
o
p
ean
Co
nfere
n
c
e o
n
C
o
mput
er Ma
th
ematic
s an
d its A
p
p
licatio
ns (H
ER
CMA-20
03),
Athens, Greec
e. 2003: 1-
14.
Evaluation Warning : The document was created with Spire.PDF for Python.