TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 8, August 201
4, pp. 6393 ~ 6402
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.525
7
6393
Re
cei
v
ed
No
vem
ber 3
0
, 2013; Re
vi
sed
April 16, 201
4; Acce
pted
May 5, 201
4
Time-Weighted Uncertain Nearest Neighbor
Collaborative Filtering Algorithm
Zheng Zhi-G
ao*
1
, Wang P
i
ng
1,2
, Sun Sheng-Li
1
1
School of Softw
a
r
e a
nd Micr
oel
ectronics, P
e
kin
g
Univ
ersit
y
,
Beiji
ng 1
0
0
260
, China
2
Nation
al Eng
i
neer
ing R
e
se
a
r
ch Center for
Soft
w
a
re Eng
i
n
eeri
ng, Peki
ng
Univers
i
t
y
,
Beiji
ng 1
0
0
871
, China
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zhengz
hi
gao
@pku.e
du.cn
A
b
st
r
a
ct
T
o
overco
me
the li
mitati
ons
of the traditi
on
al
col
l
a
borativ
e
filtering
r
e
co
mme
n
d
a
tion al
gorith
m
,
this pa
per pr
o
pose
d
a T
i
me
-W
eighte
d
Un
certain N
ear
e
s
t Neigh
bor C
o
lla
bor
ative F
i
l
t
ering Al
gor
ith
m
(TWUNCF). Ac
cordi
ng to t
he actual a
ppl
icati
on sit
uati
on of reco
mme
ndati
on system, the
author w
e
ight
e
d
the pro
duct si
mi
larity a
nd us
er simil
a
rity to ensur
e the d
a
ta vali
dity firstly, and the
n
calc
u
l
ate the si
milar
i
ties
of user a
nd pr
oduct a
nd ch
o
o
se t
he truste
d
nei
ghb
or gro
u
p
as the re
c
o
mme
n
d
ed o
b
j
e
ct ada
ptively
bas
ed
on th
e w
e
ig
ht.
Experi
m
ental
r
e
sults s
how
th
at the
alg
o
rith
m c
an
be
use
d
to i
m
pr
ove
dat
a val
i
d
i
ty accor
d
in
g
to the time attri
bute, an
d ba
la
nc
e the i
m
pact
the differe
nt gr
oups
on
the r
e
commen
datio
n
result, and
av
oid
the prob
le
ms w
h
ich ca
use
d
by
the
data sp
ars
eness. T
h
e
o
ret
i
cal a
nalys
is an
d exper
i
m
e
n
tal
de
monstrati
o
n
s
show
that th
e
alg
o
rith
m th
is
pap
er
prop
ose
d
o
u
tperfo
rm
s
e
x
i
s
ti
ng
al
go
ri
th
m
s
in
re
comm
en
da
ti
o
n
qu
al
i
t
y,
and i
m
prove th
e system's acc
u
racy an
d reco
mme
n
d
a
tion ef
ficiency.
Ke
y
w
ords
:
c
o
l
l
ab
orative
filter
ing, ti
me w
e
i
g
ht, uncert
a
in
n
e
ig
hbors,
trustw
orthy subset,
reco
mmend
ati
o
n
system
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
With the
d
e
v
elopment
a
nd p
opul
ari
z
ation
of
e-commerce,
m
any researchers
and
schola
r
s h
a
ve made the
relevant re
sea
r
ch
on t
he re
comm
end
atio
n efficien
cy and accu
ra
cy of
recomme
ndat
ion syste
m
in ord
e
r to mi
ning pote
n
tial
custo
m
ers g
r
eatly
. Many schola
r
s
hav
e
prop
osed a
variety of recommen
d
a
tion algo
ri
thm, among
which coll
aborative filterin
g
recomme
ndat
ion algo
rithm
is the mo
st wi
dely use
d
.
Curre
n
tly the resea
r
ch, based on
colla
bo
rative filtering
,
is mainly divided into two
kind
s:
use
r
-ba
s
ed
collabo
rative filtering a
nd p
r
odu
ct-ba
s
e
d
colla
borative filtering. Eithe
r
the u
s
e
r
or t
he
prod
uct, there is individu
a
l
variation, which
re
sults i
n
the re
com
m
endatio
n di
fferences. In
the
use
r
-ba
s
ed
recom
m
en
dati
on, the tra
d
i
t
ional u
s
er
-b
ase
d
collab
o
r
ative filterin
g algo
rithm
or
prod
uct
-
ba
se
d one
ha
s so
me limitation
s
, due to
th
e
inherent diffe
ren
c
e
s
am
on
g users an
d t
h
e
uncertainty of scenarios predi
ct
ion
and the vari
abilit
y of production predi
ction, mainly in t
h
e
followin
g
thre
e area
s: (1
)
Many sch
o
lars u
s
ually
use
the
k
NN
[1-3]
method
to recomme
nd an
objec
t for the target. Sinc
e the
k
NN method
choo
se
s
k
nea
re
st
neigh
bo
r
th
roug
h a ce
rtain
simila
rity co
mpari
s
o
n
, it can
pre
s
e
n
t the ch
ara
c
te
ristics of the
predi
cted ta
rget to a cert
ain
extent. Howe
ver,
k
i
s
a
common
argu
ment an
d u
s
ually
do n
o
t have the p
a
rticularity, whi
c
h
make
s it may
be not availa
ble in certai
n
scena
rio
s
. Fo
r exampl
e, an
extreme
ca
se that wh
en t
h
e
numbe
r
of th
e obj
ect’
s n
e
i
ghbo
r i
s
sma
ller tha
n
k
, the
k
NN m
e
th
ods will
re
co
mmend
seve
ral
individual
s which is q
u
ite
different from t
he object
as the object’s nei
ghb
or, makin
g
the
recomme
ndat
ion un
relia
ble
.
(2)
Cu
rre
nt recom
m
en
dati
on alg
o
rithm
often only re
commen
d
for
a
certai
n
singl
e
gro
up
of u
s
e
r
s or produ
ct
s, ign
o
ri
n
g
th
e impa
ct fo
r t
he oth
e
r
grou
ps [3
-6]. Sin
c
e
we neith
er h
a
ve deeply u
nderstan
ding
in the reco
mmend
ed in
dividual dime
nsio
n tru
s
two
r
thy
sub
s
et befo
r
e
we re
comm
end, nor
stud
y its in
fluence on re
comm
endatio
n re
su
lt, the quality
of
recomme
ndat
ion to some e
x
tent can’t meet the
need
s of people in every asp
e
ct.
(3) Tradition
al
recomme
ndat
ion algo
rithm doe
s not con
s
ide
r
the fact
that the valu
e of the data decrea
s
e
s
wi
th
time when
m
a
ke
de
ci
sion
s. Treatin
g
different
dat
a
du
ring
differe
nt
perio
ds affe
cts the
a
c
curacy
and fea
s
ibility of recom
m
en
dation to som
e
extent.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 639
3 –
6402
6394
In order to
sol
v
e the th
ree
i
s
sue
s
a
bove,
this
pap
er,
u
s
ing
the i
dea
of dynami
c
selectio
n,
on the
ba
sis of the exi
s
ting resea
r
ch,
ada
ptiv
ely choo
se
re
com
m
ende
d o
b
je
ct’s trust
w
o
r
thy
sub
s
et in
different di
men
s
ions
und
er di
fferent scen
a
r
ios
and
req
u
irem
ents
as recomme
nd
ed
can
d
idate
set
,
and p
r
opo
se a Time
We
ighted
Un
cert
ain Neigh
borhood
Coll
abo
ration Filte
r
in
g,
TWUNCF. Th
is algo
rithm,
by using
use
s
logi
stic
m
e
thod
s to wei
g
h time for score
s, distin
gu
ish
score
s
i
n
diff
erent
pe
riod
s and
fully con
s
ide
r
the
fact
that the valu
e
of
data
de
creases over time
to ensu
r
e th
e validity of the data ag
ai
nst time. As
to the data with sa
me time attribute,
it is
cal
c
ulate
d
ba
sed o
n
the si
milarity of users a
nd p
r
od
ucts, in the
meantime it determi
ne
s which
trust
w
orthy
subset of re
co
mmend
ed target as
th
e re
comm
end
ed
can
d
idate
set
and sele
cts t
he
neigh
bor
of p
r
edi
cted ta
rg
et adaptively. Experime
n
ts
sho
w
that th
is alg
o
rithm,
Time Weight
ed
Un
certai
n
Ne
ighbo
rho
od Collabo
ration Filtering, ca
n
effectively balan
ce the i
m
pact of diff
erent
grou
ps on
re
comm
end
atio
n, ha
s g
ood
p
e
rform
a
n
c
e i
n
solvin
g sco
r
e data
sparse
ness
pro
b
lem
s
,
and ta
ke
s int
o
a
c
count
th
e imp
a
ct
of t
he di
minishin
g value
of ti
me o
n
the
recom
m
en
dati
o
n
result. As a result, the p
r
o
posed alg
o
rit
h
m in this p
a
per imp
r
ove
s
the quality of recomme
ndat
ion
and ha
s go
o
d
stability to some
extent. This pa
pe
r is organi
ze
d as follo
ws: S
e
ction 2
de
scribe
s
related
wo
rk and defin
es the issue
s
to be solv
ed
in this pa
per; Section 3
details the Ti
me
Weig
hted Un
certai
n Neigh
borh
ood
Coll
aboration
Filt
ering
algo
rith
m; Section
4
desi
g
n
s
multi
p
le
experim
ents
to validate the prop
osed a
l
gorithm
an
d make
s a sim
p
le analysi
s
; finally make
a
summ
ary
.
2. Relev
a
nt Work a
nd Problem Defini
tion
2.1. Relev
a
nt Work
Currently, data sparseness
[7],
cold
start [8]
and scalability [9] are
com
m
on i
n
recomme
ndat
ion system.
To solve the
s
e thre
e is
su
es, many re
searche
r
s hav
e made a lot
of
resea
r
ch, hop
ing to improv
e them by a n
e
w algo
rithm.
For example,
Seung-T
a
e
k
Park
cre
a
ted a
new
sea
r
ch model MAD6
[10] by combining
collab
o
rative filterin
g algorith
m
with search e
n
g
ine
tools, and ap
plied it to Yahoo! [4]; Tomoharu us
ed
maximum ent
ropy pri
n
cipl
e
to predict th
ose
prod
uct
s
con
s
ume
r
s inte
re
sted
ba
sed
o
n
collabo
rativ
e
filtering
alg
o
rithm [5], a
n
d
it turn
ed
ou
t to
be a su
cce
s
s wh
en ap
pli
ed to E-co
m
m
erce sy
ste
m
; Chen [2] etc. applie
d dual collab
o
rative
filtering algo
ri
thm to search for pro
d
u
c
ts that
target
use
r
s m
a
y be intere
sted i
n
, which usi
n
g
colla
borative filtering meth
ods a
gain in
the firs
t re
sult set for a se
con
d
re
co
mmend
ation;
Gu
prop
osed a
time weig
h
t
ed re
comm
endatio
n alg
o
rithm, taki
n
g
time prop
erties [
11] i
n
to
con
s
id
eratio
n
pro
perly in
t
he process
of re
comm
en
dation;
Huan
g brought
ou
t an un
ce
rtai
n
nearest
n
e
ig
hbor re
com
m
endatio
n
a
l
gorithm whi
c
h comp
re
h
ensively con
s
ide
r
ed different
grou
ps
of users
and p
r
o
d
u
cts, b
a
lan
c
i
ng the
u
s
er
grou
p an
d the pro
d
u
c
t group [12]; Ch
en
improve
d
resource
asse
ssment de
nsity
throug
h th
e
e
s
tabli
s
hme
n
t
of k
nea
re
st
neigh
bor an
d
its
impact
set, and he defin
e
d
a new
re
commen
dation
mech
anism to calculate the sco
r
e of the
predi
ction [1
3], which all
e
viated the
data spa
r
sen
e
ss p
r
oble
m
effectively and imp
r
oved
the
quality of the recomme
ndat
ion; Liu u
s
ed
Beta distrib
u
tion to predict t
he simil
a
rity of use
r
s
ba
se
d
on tru
s
tworth
y grou
p, imp
r
oving th
e re
comm
end
ed
re
sult to a
certain
extent [14]; Jamali,
in
orde
r to
imp
r
ove the
qualit
y of re
com
m
endatio
n, ma
de a
de
ep
di
gging
in th
e t
r
ust
rel
a
tion
ship
betwe
en u
s
e
r
s, foun
d the
deep u
s
e
r
simila
rity
and
made a
re
commen
dation
by some
da
ta
mining meth
o
d
s [15], and finally improve
d
t
he re
comm
endatio
n efficiency of the system.
2.2. Problem Definition
Traditio
nal collabo
rative filtering alg
o
ri
thm finds k
neigh
bors who influen
ce
curren
t
individual m
o
st thro
ugh th
e inne
r indivi
dual inte
ra
ction bet
wee
n
the u
s
er
grou
p and th
e p
r
o
duct
grou
p
to pre
d
ict curre
n
t
i
ndividual pro
perty.
Bu
t wi
th the in
crea
sing
u
s
e
of recom
m
en
dati
o
n
system a
nd
the increa
sin
g
com
p
lexity of
the environm
ent, this method of
intercepting
k
neigh
bors tu
rns to
be
pa
rtial. This on
e-side
dne
ss
i
s
mainly ma
nifested
in t
w
o
asp
e
ct
s. On
e
is
that con
s
ide
r
i
ng se
parately the impact o
f
a par
ticula
r
grou
p and ig
norin
g the im
pact of anoth
e
r
grou
p
whi
c
h
itself is un
re
aso
nabl
e. Th
e othe
r o
ne i
s
that
wh
en t
he d
a
ta of in
dividual
s in t
h
e
grou
p is
spa
r
se, the num
b
e
r of neigh
bo
rs in the
clu
s
ter may be le
ss than the val
ue of
k
in
k
NN
,
the ba
sic
re
commen
dation
algo
rithm if
simply
recom
m
end based on
u
s
e
r
s or p
r
odu
cts.
In
th
is
ca
se, users o
r
pro
d
u
c
ts
with low
simila
rity have to
be adde
d into th
e trainin
g
set,
thus lea
d
ing
to
a sha
r
p de
cli
ne in the accuracy of the
algorithm
. And its re
sult is often inaccurate or ev
en
wro
ng.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
-Weig
hted Un
ce
rtain
Nea
r
e
s
t Neig
hbor
Colla
bor
ative Filteri
n
g
Algorithm
(ZHENG Z
h
i-g
a
o
)
6395
For
exampl
e, the m
e
thod
s cu
rrently u
s
e
d
to
pr
ed
ic
t
ho
w
muc
h
a
pe
r
s
o
n
lik
es an
ite
m
is
User
Based Colla
borative Filteri
ng, UB
CF and
Item Based
Collab
o
rative
Filte
r
i
ng,
IBCF.
T
h
ough
these
metho
d
s
to
som
e
ext
ent are
cap
a
b
le of
re
com
m
endin
g
, the
con
s
trai
nts
of reality ma
ke
its
ac
cur
a
cy
f
a
r
more t
h
a
n
sat
i
sf
act
o
ry
.
For
inst
an
ce,
when a user is int
e
rested in an item while few
peopl
e did a
nd mad
e
little evaluation,
the ac
cu
ra
cy of recom
m
endatio
n ba
sed on
UBCF
is
relatively low.
Due to the compl
e
xity of the r
eality,
we nee
ds to
con
s
id
er every aspect of the
use
r
s an
d th
e produ
cts
a
daptively sel
e
cts in u
s
e
r
grou
p a
nd p
r
odu
ct group,
but not
sim
p
ly
con
s
id
er on
e
aspe
ct and i
gnore the oth
e
rs
whe
n
ma
king a recom
m
endatio
n. Acco
rdi
ng to the
need
s of
actu
al situatio
n, we ca
n dyn
a
mi
cally
sele
ct th
e nea
re
st nei
ghbo
r a
nd th
e neig
hbo
rho
o
d
factor, pi
ck n
e
ighb
ors p
r
o
perly o
u
t of u
s
er g
r
ou
ps
a
nd p
r
od
uct
group
s a
nd al
so re
co
mmen
d
fo
r
the curre
n
t object.
Even the
nu
mber of th
e
use
r
i
n
the
u
s
er
cl
u
s
ter or the
numb
e
r
of the
pro
d
u
c
t in the
prod
uct
clu
s
ter satisfie
s the lowe
st valu
e of k in
kNN,
time incon
s
i
s
tency may also exist. Since
a
use
r
’s i
n
tere
st may chang
e
with time, the score t
hat a use
r
give
s to the sam
e
p
r
oje
c
t may vary
with time to
o
.
Ho
wever, t
r
aditional
algo
rithm tre
a
ts
a
user’
s
score
s
e
qually in
f
i
nding
a u
s
e
r
’s
nearest n
e
ig
hbor witho
u
t taking th
e
variation
with
time of the
use
r
’
s
interest into a
c
co
unt,
resulting in th
e cal
c
ulate
d
may not be the neig
hbo
r
grou
p the u
s
er re
ally interests. Th
e poi
nt is
the accuracy
of the kNN al
gorithm de
pe
nds on
h
o
w
much the
sel
e
cted nei
ghb
ors m
a
tch wit
h
the
target
use
r
s,
whi
c
h i
s
one
of the i
m
po
rtant re
aso
n
s
why the
a
c
cu
racy
of tra
d
itional
algo
rith
m
sho
u
ld
be im
proved.
For
example, a
u
s
er A
wa
s int
e
re
sted i
n
a
c
tion movie
s
durin
g a
cert
ain
perio
d of the
past
and
scored
highly
on such film
s while
anot
her
user B i
s
inte
re
sted i
n
it
curre
n
tly
and score
s
highly on
the sam
e
f
ilms.
Acco
rd
i
ng to the
cal
c
ulation p
r
in
ci
ple of
simila
rity,
the two u
s
ers sh
ould
be
each othe
r’s
neigh
bors. But in fact, it’s obviou
s
ly
unre
a
sona
ble
to
recomme
nd
based o
n
two
use
r
s’ intere
sts in
differe
n
t
time. Firs
t of all, the c
u
rrent interes
t
of A
may not be a
s
same a
s
B’
s. Seco
ndly, influen
ced by
so
cial po
pula
r
ity, what A like
s
at that time
may not
be f
a
vored
by B
cu
rrently. Usually,
sea
r
ching th
e
simi
larity amo
ng
the sco
r
e
s
th
at
different u
s
e
r
s give t
o
the
same
item
wi
thin the
sam
e
or si
milar p
e
riod
of time
can
en
su
re t
h
e
effectiveness of
the
selected neighbor. T
able 1 illust
rates it with an example.
Table 1. Simple Example
User
Project
I
1
I
2
I
3
I
4
I
5
I
6
u
1
(T
1
)
4 3
4 3
3
4
u
2
(T
4
)
3 4
2 3
4
2
u
3
(T
1
)
3 4
4 3
3
4
u
4
(T2
)
4 3
2 4
3
2
Table
1
sho
w
s the
sco
r
e
s
that 4 u
s
e
r
s g
i
ve to 6 p
r
oje
c
ts i
n
3
peri
o
ds
whe
r
e
T1
and T
2
are si
milar p
e
r
iod of time, T1 and T4 a
r
e
far away fro
m
each oth
e
r.
Acco
rdi
ng to the assumpti
on above,
u
2
、
u
3
and
u
4
constitute the nearest nei
gh
bor set of
u
1
when re
comm
end with 3 neigh
bor u
s
ers
by the traditional colla
b
o
rative filtering
recomme
ndat
ion alg
o
ri
thm, and
the
similarity satisfies th
e conditio
n
12
1
3
1
4
,,
,
s
i
m
uu
s
i
m
u
u
s
i
m
uu
. However, a
c
cordi
ng to th
e data given i
n
Table
1, the time
gap is l
a
rg
e b
e
twee
n the time wh
en u
s
e
r
u
1
and
use
r
u
2
each sco
r
es the
sam
e
proje
c
ts. A
s
the
analyses
abo
ve, it is un
rea
s
on
able to
predict the
interest of u
s
e
r
u
1
based
on th
e sco
r
e d
a
ta
of
use
r
u
2
in a certai
n peri
o
d of the last and give a re
conm
end
atio
n to
u
1
.
If
tak
i
ng the time into
con
s
id
eratio
n
,
only
u
3
and
u
4
is the neig
hbor of
u
1
, whic
h is
more c
o
ns
is
tent wit
h
the fac
t.
3. Time-Wei
ghted
Unc
e
r
t
ain Ne
ares
t
Neighbo
r Co
llaborativ
e F
iltering Algo
rithm
3.1. Time-We
ighted Near
e
s
t Neig
hbor
Without limit
s to the pe
rio
d
of time
wh
en se
le
cting t
he k nei
ghbo
rs
of the u
s
e
r
or the
item, it is e
a
sy to ta
ke
outdated
d
a
ta as
neig
hbors i
n
to consi
deration
(Li
k
e m
a
ke
a
recomme
ndat
ion ba
sed on
the thing that a user inte
re
sted 20 yea
r
s ago), whi
c
h t
u
rn
s out to be
unsatisfa
ctory. In the traditional
colla
borative f
iltering alg
o
rithm
,
not disting
u
i
shin
g the d
a
t
a’s
time effec
t
iveness
,
to
s
o
me extent, affec
t
s
the ac
curacy of th
e re
sult. Ta
king
the influe
nce t
hat
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 639
3 –
6402
6396
the time h
a
s
on the
targ
et
grou
p
com
p
rehen
sively
a
nd mo
difying
the simil
a
rity
by wei
ghing
the
time can avo
i
d the low a
c
curacy of the
reco
mmen
d
a
tion re
sult caused by time inco
nsi
s
ten
c
y
effec
t
ively.
As a re
sult, this pa
pe
r pro
poses a Ti
m
e
Weig
hted
Selected
Nei
ghbo
r metho
d
. Before
sele
cting the
neigh
bors, time-wei
g
h
s a
nd modify the sco
re
s with
log
i
stic
function. In this pap
er,
it use
s
the P
earson
Co
rre
l
ation Metho
d
to cal
c
ulat
e
the simila
rity. Firstly, wei
g
h the sco
r
e
s
o
f
different use
r
s an
d different peri
o
d
s
of time with
lo
gistic
function, replace
,
uc
r
with
,
,
uc
uc
r
l
o
g
istic
r
, ma
k
i
ng
ea
ch
sc
or
e
h
a
s
its
time
w
e
ight. Sin
c
e
th
e la
te
s
t
s
c
o
r
e re
p
r
es
en
ts
a
use
r
’s inte
re
st most, the
current
weig
ht sh
ould
be
la
rge
r
tha
n
the
last
one
an
d the
sco
r
e
s
of
different time sho
u
ld be di
stinguished.
Here is
the
log
i
stic
f
unc
tion.
,
,
1
()
1
uj
uj
t
logistic
t
e
(
1
)
Among
the
m
,
11
t
,
,
0(
)
1
uj
l
o
gi
stic
t
.
,
()
uj
l
o
gi
stic
t
is a
monotoni
c i
n
cre
a
si
ng
function, the weig
ht increa
se
s with time
t increa
sing,
and the value is alway
s
within (0,1). Th
is
pape
r firstly converts th
e range
of time to [-1,1]
by using
a sta
nda
rdiz
ed conve
r
sion
m
e
thod. By
doing thi
s
, the weig
ht’s variation with tim
e
is al
mo
st lin
ear, an
d the
cha
nge of u
s
er’s i
n
tere
st can
be dete
c
ted i
n
tuitively
[16].
The most important step in collaborative filter
ing alg
o
rithm is the formation of the
target
use
r
’s
neig
h
b
o
rs.
He
re it u
s
e
s
the ne
arest nei
ghbo
r
whi
c
h is
used
in the co
mp
uting of Pearson
correl
ation. The Pearso
n relevant simila
ri
ty based o
n
time-weighte
d
of the use
r
s is:
,,
,
,
,,
,
,
22
((
)
)
(
(
)
)
(,
)
((
)
)
(
(
)
)
ac
ac
a
b
c
b
c
b
U
ac
ac
b
c
b
c
UU
i
UU
U
U
U
U
cI
ab
UU
U
U
cI
cI
l
ogi
s
t
i
c
l
ogi
s
t
i
c
l
ogi
s
t
i
c
l
ogi
s
t
i
c
rt
r
r
t
r
si
m
U
U
rt
r
t
(2)
In the above formula,
(,
)
ab
sim
U
U
is the simila
rity betwe
en the target u
s
e
r
a
U
an
d its
nearest
neig
h
bor
b
U
,
U
I
p
r
es
en
ts
th
e pr
o
j
ec
t
s
e
t th
a
t
b
o
t
h
u
s
er
a
U
an
d
b
U
scor
e
s
,
the sc
ore
use
r
a
U
gives t
o
proje
c
t c i
s
, the ea
ch a
v
erage
sco
r
e
that user
a
U
an
d
us
er
b
U
give to a
proje
c
t i
s
a
U
r
and
b
U
r
. Assuming
that
a
U
is the
ta
rget
user,
we
ca
n p
r
e
d
ict t
he
score
that
a
U
giv
e
s wit
h
t
h
e
sco
re
s
b
U
gives to any item
()
U
j
jI
.
Here is
the formula:
,
1
,
1
[(
,
)
(
)
]
[(
,
)
]
ac
a
b
n
ab
U
U
i
uj
U
n
ab
i
s
im
U
U
r
r
Pr
sim
U
U
(
3
)
Here it
calcul
ates th
e si
milarity by time
-weig
h
ted Pe
a
r
so
n
Correl
ation Fu
nctio
n
,
and the
simila
rity calculation of the
use
r
s a
nd th
e one of
the
proje
c
ts i
s
ca
rrie
d
out at the sam
e
time. It
predi
cts
the value
that
a
use
r
score
s
other pr
oj
ect
s
, an
d then
take
s th
ose items
whi
c
h
d
on’t
belon
g to
U
I
into the re
com
m
endatio
n se
t in descen
d
i
ng mod
e
. And finally sele
ct the prope
r
proje
c
t from t
he re
comm
en
dation set to make a
re
co
mmend
ation.
3.2. Impro
v
e
d
Similarit
y
Computing
The simila
rity
amo
ng users
i
s
wei
ghed
with
th
e
scores th
at different u
s
e
r
s giv
e
to th
e
same
p
r
oje
c
t
s
o
r
th
e
sam
e
items. If th
e
numbe
r
of
th
e same
proje
c
ts
or item
s t
he u
s
e
r
s
score is
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
-Weig
hted Un
ce
rtain
Nea
r
e
s
t Neig
hbor
Colla
bor
ative Filteri
n
g
Algorithm
(ZHENG Z
h
i-g
a
o
)
6397
so
sm
all tha
t
it is
una
bl
e to m
eet t
he mi
nimum
num
ber of
neigh
bor,
the
accu
ra
cy of
its
recomme
ndat
ion re
sult wi
ll decrea
s
e.
In orde
r to avoid this
occa
sion
al influen
ce, so
me
resea
r
chers
alrea
d
y do some re
se
arch on it. Herlo
c
ker et
c. improves the
si
milarity cal
c
ul
ation
with an in
cre
a
se
d wei
ght of the releva
nce; Ma
et
c. points o
u
t its spe
c
ific settings in d
o
cum
ent
[17]
. And this
pape
r
control
s
it by
setting
a threshold.
Suppo
sing
th
at
'
ab
I
UU
presents t
h
e
items that b
o
t
h use
r
a
U
and use
r
b
U
have
scored,
we a
d
d
the p
r
op
ort
i
on of differe
nt use
r
s
score in the
same or
simila
r time perio
d to improve the
similarity am
ong u
s
ers.
mi
n
(
|
'
|
,
)
'(
,
)
(
,
)
ab
a
b
I
s
im
U
U
s
i
m
U
U
(
4
)
Among them
is the thresh
old, and a
c
co
rding to it
s de
finition its ma
ximum value i
s
the
numbe
r of th
e proj
ect
s
th
at the use
r
s
both sco
r
e
s
. So
mi
n
(
|
'
|
,
)
1
I
, and the range of the
improve
d
use
r
simil
a
rity
'(
,
)
ab
sim
U
U
is still [0,1]. When the n
u
m
ber of the
sa
me proje
c
ts t
he
use
r
s
sco
r
e i
s
large
r
than
its thre
shol
d,
'(
,
)
(
,
)
ab
a
b
si
m
U
U
s
im
U
U
. In
th
e
s
a
me w
a
y, w
hen
the num
ber o
f
the sam
e
p
r
oject
s
the
users sco
r
e
i
s
small, the si
mil
a
rity amon
g u
s
ers
sh
ould
b
e
decrea
s
e
d
to improve its in
fluence on th
e recomme
nd
ation accu
ra
cy.
3.3. Uncer
tai
n
Near
est
Ne
ighbor Trus
tw
o
r
thy
Subset
Curre
n
tly the
use
r
-ba
s
ed
o
r
p
r
od
uct-ba
sed
co
llab
o
rative filtering
al
gorithm
is co
mmon i
n
recomme
ndat
ion
system.
Howeve
r d
ue t
o
the va
riabili
ty of the user’
s
d
e
man
d
s o
n
the q
uality
of
recomme
ndat
ion and th
e reality, the recommen
dation
always tu
rn
s out to be un
able to me
et the
use
r
’s
need
s if only using
one metho
d
s
espe
cially
whe
n
the dat
a is sparse
n
e
ss. That ho
w to
con
s
id
er several facto
r
s
comprehe
nsiv
ely and how to
weigh them
automatically
is what we n
eed
to solve. Aiming at the ch
ara
c
teri
stic of
a
reco
mme
ndation
syste
m
makin
g
de
cisi
on with th
e
nearest nei
gh
bors.
This p
ape
r o
p
timize
s the
sele
ction of recom
m
en
dati
on set, an
d weig
hs b
e
tween the
use
r
an
d the
prod
uct a
dap
tively to avoid too mu
ch hu
man involve
m
ent whi
c
h
result
s in redu
ced
flexibility of re
co
mm
endation system.
In docume
n
t [12], it propo
se
s a m
e
thod
to we
ig
h the
use
r
’
s
simil
a
rity and the
p
r
odu
ct’
s
by setting two simila
rity threshold
and
, and dynami
c
ally sel
e
ct th
e pro
per
neig
hbor
of the
recomme
ndat
ion target in use
r
’s g
r
ou
p and produ
ct
’s group b
e
fore
sele
cting its neigh
bors. Th
i
s
method, to some extent, can ove
r
com
e
t
he sh
ort
c
oming of tra
d
itional alg
o
rithm
k
NN
a
nd
dynamically sele
ct the neighb
or, but it needs 2
thre
shol
ds to
weigh it. Howeve
r, setting a
threshold
ha
s som
e
influe
n
c
e o
n
nei
ghb
or
sele
cting,
and its calcul
ation is rel
e
vantly com
p
le
x. If
the threshol
d is set too
big, the num
ber of
recommendation
group
will decrease and the
recomme
ndat
ion will lack of generality; if
the thresh
old is set too
small, it is easy to bring i
n
neigh
bor with
low simil
a
rity
whi
c
h
affect
s the a
c
curacy
of the
re
com
m
endatio
n. T
herefo
r
e, i
n
t
he
absen
ce of p
r
ior
kno
w
led
g
e
, this meth
o
d
is
still
not
well eno
ugh to
sele
ct the u
s
er’s neig
hbo
r.
In
orde
r to ove
r
co
me the
shortcomin
g in docum
e
n
t [12], this paper
brin
gs
in a Ha
rmo
n
ic
para
m
eters to balan
ce the
user-b
ased
method an
d the pro
d
u
c
t-b
a
se
d method.
Here, we n
o
te the gro
u
p
of the predicte
d
sco
r
e that use
r
a
U
to item
j
I
as
12
'
'(
)
{
,
,
,
}
m
aa
a
a
SU
U
U
U
, and note th
e size of the
grou
p a
s
|'
(
)
|
'
a
SU
m
. The item gro
up
that
use
r
a
U
has scored in the
neighbo
r of item
j
I
is noted as
12
'
()
{
,
,
,
}
n
jj
j
j
SI
I
I
I
, and
|(
)
|
'
j
SI
n
. The
co
mput
ing p
r
o
c
e
dure of
'
m
an
d
'
n
is
relatively e
a
sy so
that it
can b
e
done
off-line. Thi
s
pape
r int
r
od
u
c
e
s
n
e
igh
bor facto
r
s
and
1
as the
bal
an
ce fa
cto
r
of
the u
s
e
r
grou
p a
nd th
at of the
item
group
sepa
rately, and
ad
justs the val
u
e of the
nei
g
hbor fa
ctor
with
harm
oni
c parameters.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 639
3 –
6402
6398
'
''
0
''
0
0.
5
m
mn
mn
m
mn
mn
其他
'
''
0
''
0
1
0.
5
n
mn
mn
n
mn
mn
其他
Among them,
is the harm
onic p
a
ra
met
e
r. In the followin
g
, we take
for exampl
e to
analyze h
o
w
adj
ust
s
th
e nei
ghb
or f
a
ctor.
When
''
0
mn
,
ha
s th
e f
o
llowin
g
fo
ur
possibl
e values:
Whe
n
'0
m
as well as
'0
n
, it means m
a
ki
ng
recomme
ndat
ions
by item-based
colla
borative filtering meth
od. In this
ca
se,
0
, the value of
has
nothing to do with the
value of
.
Whe
n
'
'
n
m
,
10
.
5
. In this ca
se, it means h
a
lf of th
e recomme
n
dation se
t
half come
s from item-ba
s
e
d
recomme
nd
ation set an
d half use
r
-ba
s
ed re
comm
en
dation set.
Whe
n
'
(,
)
'
n
m
,
is an
increa
sing f
unctio
n
, and
its ran
ge is
(0.5,1).
1
is an
decrea
s
in
g f
unctio
n
, its
range
is
(0,0.
5
). It mea
n
s that with th
e in
cre
a
si
ng
value of
, the
proje
c
ts in
re
comm
end
atio
n set come
s
from the
u
s
er gro
up m
o
re. An extrem
e
ca
se i
s
'0
n
and
1
, which
mean
s the m
e
thod is u
s
e
r
-bas
ed collab
o
rative filterin
g method.
Whe
n
'
[0
,
)
'
n
m
,
is an
decrea
s
ing f
unctio
n
, and
its rang
e is (0,0.5). It means the
recomme
ndat
ion set grad
ually tends t
o
the item grou
p. Whe
n
0
, it means making
recomme
ndat
ion with item-based collab
o
rative filterin
g method.
In ord
e
r to b
a
l
ance the
imp
a
ct of
user group
and
item
grou
p o
n
diff
erent
dime
nsi
on, this
pape
r intro
d
u
ce
s a h
a
rmonic
para
m
eters
to
automatical
ly adjust th
e so
urce o
f
recomme
ndat
ion
set, avoid
the lo
w
quali
t
y from reco
mmendi
ng
by sin
g
le
gro
u
p
.
The im
proved
method in thi
s
pa
per
re
du
ce
s the u
s
e
r
’
s
interve
n
tion
to the algo
rit
h
m by re
du
ci
ng the n
u
mb
er
and the difficulty of the thresh
old
s
the u
s
er n
eed
s to assign to. Co
mpared to the threshold
s
i
n
document [1
2], the on
e
in this pa
pe
r is mo
re
consi
s
tent
wit
h
u
s
er ha
bits, redu
cing
the
compl
e
xity of the
use
r
op
eration.
In d
o
c
ume
n
t [12],
and
is
se
nsitive to the
selectio
n of
neigh
bors, re
sulting in big
influence o
n
reco
mmen
d
a
tion re
sult, while this p
a
per introdu
ce
s a
harm
oni
c p
a
rameters to
co
ntrol the
thre
shol
d, re
du
ci
ng the
sen
s
itivity of neighb
or
sele
ction
a
nd
ensurin
g the
accuracy
of recom
m
en
dati
on result.
In the othe
r ha
nd, the th
reshold’
s setting
in
this pa
pe
r is
relatively ea
sy, and the
be
st thre
sh
old
can be
obtai
n
ed after multi
p
le test
s, whi
c
h
also e
n
surin
g
the algorithm
’s accu
ra
cy and red
u
ci
ng h
u
man’
s interv
ention.
3.4. Time-We
ighted Uncer
tain Ne
ares
t
Neighbo
r Co
llaborativ
e F
iltering Algo
rithm
Acco
rdi
ng to
the an
alysi
s
above, thi
s
pape
r
fully consi
ders th
e
time attribute
of the
score, a
nd
weigh
s
the u
s
er-ba
s
ed
sco
r
e an
d the it
em-b
ased on
e ba
sed o
n
that. It takes i
n
to
accou
n
t man
y
aspect
s
, an
d prop
oses a
Time Weight
ed Un
ce
rtain Neig
hbo
rho
o
d
Collab
o
rati
on
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
-Weig
hted Un
ce
rtain
Nea
r
e
s
t Neig
hbor
Colla
bor
ative Filteri
n
g
Algorithm
(ZHENG Z
h
i-g
a
o
)
6399
Filtering al
go
rithm(T
W
UNCF) to p
r
e
d
ict the
use
r
’s
score. If the average
score
of use
r
,
ax
UU
on produ
cts i
s
repres
ente
d
se
parately by
,
ax
R
R
and the
averag
e sco
r
e of a kno
w
n
use
r
o
n
prod
uct
s
,
jy
I
I
is repre
s
e
n
ted separately
by
,
jy
R
R
, then th
e T
W
UNCF
propo
sed in
this
pap
er
can b
e
rep
r
e
s
ented by the formul
a belo
w
.
,
,
()
()
,
()
(
)
'(
,
)
(
,
)
'(
,
)
(
,
)
()
(
1
)
(
)
'(
,
)
'(
,
)
yj
xa
xa
yj
jy
a
y
y
ax
x
j
x
IS
I
US
U
aj
a
j
ax
j
y
US
U
I
S
I
s
im
I
I
R
R
sim
U
U
R
R
RR
R
sim
U
U
sim
I
I
Based
on un
certai
n scen
e
s
, the algo
rithm TWUNCF
firstly bring
s
in a time variable by
log
i
stic
function, wei
gh time again
s
t the similarit
y
between th
e use
r
gro
up and the item grou
p to
disting
u
ish the time effe
ctivene
ss
of the sim
ila
rit
y
and secon
d
ly on this
basi
s
con
s
id
ers
comp
re
hen
si
vely the similarity of the user a
nd t
he item, controls the neigh
bo
r factor of different
grou
ps by
ha
rmoni
c fu
ncti
on in
dire
ctly
and th
en
co
n
s
ide
r
s the
im
pact
of the
u
s
er a
nd th
e it
em,
and ultimatel
y
produ
ce the
reco
mmen
d
e
d
result set.
T
h
e
r
e
f
o
r
e
,
the
r
e
a
r
e
2
s
t
ep
s
.
Ste
p
1
,
se
le
c
t
th
e
tr
us
tw
o
r
th
y
s
u
b
s
et o
f
th
e
r
e
c
o
mme
n
ded
target; step
2, make a
recom
m
en
dati
on ba
sed
on
the sele
cte
d
re
comm
en
dation set a
n
d
produce the final result. The
descri
p
tion
of the algorithm
is illustrated in Figure 1.
Figure 1. TWUNCF Algo
rithm
In TWUNCF
algorithm, the firs
t thing to
do is
to
dete
r
mine the
score matrix of u
s
ers and
items. Thi
s
could b
e
do
ne
off line to sav
e
time compl
e
xity of the algorithm. And
it is
22
()
Os
t
in
the wo
rst ca
se. Since bot
h
|'
(
)
|
'
a
SU
m
and
|(
)
|
'
j
SI
n
are consta
nt, the time com
p
lexi
ty o
f
comp
uting th
e neigh
bo
r sand the trust
w
orthy
sub
s
et is
('
'
)
(
1
)
Om
n
m
n
O
. In the
bes
t
condition, the time com
p
lexity is
(1)
O
, effectively avoiding the calculat
ing proble
m
cau
s
e
d
by
data
spa
r
sen
e
ss a
n
d
dat
a a
c
cumul
a
tion. And
com
pare
d
to
the
DS
N al
go
rithm p
r
op
osed
in
document [1
2], the algo
ri
thm paramet
er in thi
s
pa
per i
s
le
ss,
whi
c
h
redu
ce
s the im
pa
ct of
human o
p
e
r
a
t
ion on neig
h
bor sele
ction
and ma
ke
s it more
con
c
i
s
e
and und
ersta
ndabl
e.
TW
UN
CF Alg
o
rithm
Input: the target use
r
a
U
, the
item waited to be scored
j
I
, harm
oni
c
parameters
Output: the score
,
aj
R
that the predi
cted u
s
e
r
a
U
gives
to the item
j
I
Step 1 Separately calculat
e the use
r
si
milarity ma
tri
x
and the item similarity m
a
trix base
d
o
n
the score m
a
trix
()
Rs
t
, and save
the two matri
x
separately
)
(
_
,
A
r
r
UsrSim
s
s
and
_
,
A
rr
Ite
m
Sim
t
t
;
Step 2 Weigh time with logis
t
ic
func
tion;
Step 3 Determine the validity of the time;
Step 4 Calcul
ate
|'
(
)
|
'
a
SU
m
and
|(
)
|
'
j
SI
n
based on the u
s
e
r
’s
sco
re a
nd
the item’s
score, an
d ca
lculate the tru
s
two
r
thy sub
s
et of
'(
)
j
SI
;
Step 5 Determine a prope
r harm
oni
c pa
ramete
r
;
Step 6 Calcul
ate the value of
;
Step 7 Calcul
ate the predi
cted score
,
aj
R
that user
a
U
should
give to item
j
I
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 639
3 –
6402
6400
4. Simulation and Analy
s
is
In the following, we testify the effectivenes
s and a
ccura
cy of TWUNCF alg
o
rit
h
m with
simulatio
n
s
and explo
r
in
g descri
be the adapta
b
ili
ty
to
differen
t
size of data by using the
prop
osed T
W
UNCF alg
o
rit
h
m again
s
t to that by using
k
NN method
to verify
the prop
osed ide
a
of dynamic
sele
cted n
e
ighbo
r is correct. Mea
n
while, we try to analyze t
hat wheth
e
r the
harm
oni
c parameters’ setting ca
n lead
to a better recom
m
en
ded
result from t
he trust
w
o
r
th
y
sub
s
et. The
s
e are the two
asp
e
ct
s to be
verified in this se
ction.
Similar to th
e test
meth
ods of oth
e
r re
comm
end
ation al
gorith
m
, this p
a
p
e
r u
s
e
s
MovieLen
s d
a
taset provid
ed by Grou
pl
ens to sim
u
la
te. MovieLen
s data
s
et co
n
t
ains 10 records.
943 u
s
e
r
s
score o
n
16
82
movies fo
r five levels in tot
a
l and th
e ra
nge of the
score i
s
1
~
5. 1
point
pre
s
ent
s “p
o
o
r”, 5 is
“pe
r
f
e
ct”
,
a
nd oth
e
rs m
ean
s m
i
ddle value.
They pre
s
e
n
t use
r
intere
st
in
film in varying degree
s. The hardw
are
environ
ment of the experi
m
ent
is Intel (R) Co
re (TM
)
i5-
2520
0 2.5G
Hz
qua
d-co
re 64
-bit CP
U an
d 4
G
B of memo
ry, the software environme
n
t i
s
Wind
ows 7
64bit ope
rati
ng sy
stem (profe
ssi
onal
) and all the
cod
e
s a
r
e i
m
pleme
n
ted
with
Java(64bit JDK) and Matla
b201
2.
The de
nsity o
f
the score m
a
trix
of the user an
d the ite
m
is
100
00
0
1.
6
3
%
94
3
1
68
2
, which
mean
s
the matrix is
a spa
r
se mat
r
ix. We divide
the 943
u
s
ers from d
a
taset into 3 grou
ps to test; ea
ch
sep
a
rately ha
s 100 u
s
e
r
s,
200 u
s
er
s, 3
00 users. We
sele
ct 70%
sampl
e
data
from the wh
o
l
e
dataset as th
e train set, th
e other 30%
as the test se
t to compare and verify.
4.1. Compari
s
on of Dy
na
mic Selection Method
In the exp
e
ri
ments,
we
n
eed to
sepa
rately
verify the meth
od t
o
sele
ct tru
s
tworthy
sub
s
et an
d the one to re
comm
end. Fi
rst of all, we
compa
r
e the
DSN metho
d
– a metho
d
to
sele
ct tru
s
tworthy su
bset
- with the kNN met
hod to
test whethe
r the prop
ose
d
DSN meth
od
coul
d succe
s
sfully pick o
u
t
the relativel
y
good
n
e
igh
bor o
b
je
cts,
and p
r
ep
are
for the follo
wi
ng
recomme
ndin
g
. We
set
k as th
e ab
scissa an
d
co
mpare the
p
e
rform
a
n
c
e
o
f
the 2 different
method
s i
n
d
i
fferent n
e
igh
bors, a
nd it
s
rang
e
i
s
1,2,4,8,10…6
0
.
Measure it b
y
the MAE(M
e
a
n
Absolute Erro
r). The results are
sho
w
e
d
in Figure
s
2-4.
Figure 2. Co
mpari
s
o
n
of kNN Meth
od a
nd
DSN Meth
od
(100
Use
r
s)
Figure 3. Co
mpari
s
o
n
of kNN Meth
od a
nd
DSN Meth
od
(200
Use
r
s)
Figure 4. Co
mpari
s
o
n
of kNN Meth
od a
nd DSN M
e
th
od (30
0
Use
r
s)
0.
7
0.
75
0.
8
0.
85
0.
9
0.
95
12345
678
1
0
2
0
3
0
4
0
6
0
MAE
K
100u_kN
N
100u_DSN
0
0.
2
0.
4
0.
6
0.
8
1
1
234
567
8
1
0
2
0
3
0
4
0
6
0
MAE
K
200u_kN
N
200u_DSn
0
0.
2
0.
4
0.
6
0.
8
1
1.
2
1
2
3
4
5
6
7
8
10
20
30
40
60
MAE
K
300u_kN
N
300u_DSn
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Tim
e
-Weig
hted Un
ce
rtain
Nea
r
e
s
t Neig
hbor
Colla
bor
ative Filteri
n
g
Algorithm
(ZHENG Z
h
i-g
a
o
)
6401
At first, comp
are the
experi
m
ental re
sult
s ho
ri
zontally. We can
see,
when th
e nu
mber
of
use
r
is 1
00,
kN
N ha
s b
e
s
t
result
s wh
e
n
k
=
7. Howe
ver DSN
ha
s a better re
sult in the sa
me
condition than kNN, and when k takes othe
r val
ues,
DSN i
s
still able to achi
eve bet
ter
perfo
rman
ce.
Compa
r
in
g the situatio
n of 200 user
s to that of 300 use
r
s, the
DSN ha
s bet
ter
perfo
rman
ce
than the kNN
unde
r the sa
me con
d
ition
s
.
Secon
d
ly, co
mpare the
re
sults of e
a
ch
grou
p
lo
ngitu
dinally. It’s o
b
vious that th
e big
g
e
r
the train
set i
s
, the m
o
re f
a
vorabl
e findi
ng the
targ
et user’
s
trust
w
orthy
su
bse
t
and m
a
ki
ng
a
recomme
ndat
ion i
s
. Th
rou
gh the
an
alysis an
d
com
pari
s
on,
we
can
find th
at DS
N h
a
s b
e
tte
r
stability and a
c
cura
cy than
kNN un
der th
e same
con
d
i
t
ions.
The experi
m
ent sho
w
s t
he DSN me
thod pr
o
p
o
s
ed in this paper i
s
of positive
signifi
can
c
e f
o
r improving
the method
for deter
min
i
ng the trust
subg
rou
p
. Therefo
r
e, in the
followin
g
ex
perim
ents,
we sel
e
ct the
target
o
b
je
ct’s trust
w
ort
h
y sub
s
et
with DSN
wh
en
recommending for UBCF and IBCF.
4.2. Compa
r
ativ
e Experiments
bet
w
een
the
Tra
d
itional Rec
o
mmendatio
n
Algori
t
hm
and
the T
W
UNCF
Algorithm
in this Paper
The expe
rim
ent com
p
a
r
e
s
the traditio
nal co
llabo
ra
tive filtering algorith
m
UB
CF an
d
IBCF
to
the TWUNCF alg
o
rithm pro
p
o
s
ed
in
th
i
s
p
a
per. T
he a
b
scissa
indi
cate
s the
num
ber of
the predi
cted
target item’s
neigh
bor, an
d
the ordinate
use
s
MAE as the metric.
Figure 5. Co
mpari
s
o
n
of IBCF
、
UB
CF and
TW
UN
CF (
1
0
0
U
s
er
s
)
Figure 6. Co
mpari
s
o
n
of IBCF, UBCF a
nd
TW
UN
CF (
2
0
0
U
s
er
s
)
Figure 7. Co
mpari
s
o
n
of IBCF,
UBCF a
nd TW
U
NCF
(300
U
s
e
r
s
)
Comp
ari
ng a
nd
a
nalyzi
ng the
three experime
n
ts abo
ve,
we ca
n
fi
nd
that TWUNCF
can
obtain
s
the smaller valu
e
of MAE than IBCF and
UB
CF un
de
r the
same
co
nditi
ons
and it ha
s
relatively
bett
e
r re
comm
en
dation wh
en comp
ari
ng
ea
ch
experi
m
en
t hori
z
ontally.
Com
p
a
r
ing t
he
three expe
ri
ments verti
c
a
lly, it
is obvious that the q
uality of
reco
mmend
ation i
n
crea
se
s wit
h
the
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
0.
7
0.
8
0.
9
1
1
02
0
3
04
05
0
6
07
08
09
0
1
0
0
MAE
Users
100u_TB
U
N
C
F
100u_UB
C
F
100u_IB
C
F
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
0.
7
0.
8
0.
9
1
1
02
03
0
4
05
06
0
7
08
09
0
1
0
0
MAE
Users
200u_TB
U
N
C
F
200u_UB
C
f
200u_IB
C
F
0
0.
1
0.
2
0.
3
0.
4
0.
5
0.
6
0.
7
0.
8
0.
9
1
1
02
03
0
4
05
06
07
08
09
0
1
0
0
MAE
Users
300u_TB
U
N
C
F
300u_UB
C
F
300u_IB
C
F
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 8, August 2014: 639
3 –
6402
6402
increa
sing
n
u
mbe
r
of the tru
s
twort
h
y su
b
s
et. Analyzing al
l
the
expe
riments abov
e
comp
re
hen
si
v
e
ly
,
we c
a
n
com
e
t
o
a
n
con
c
lu
si
o
n
th
at TWUNCF
has better p
e
rform
a
n
c
e t
hat
UBCF a
nd IBCF und
er the
same
con
d
itions.
5. Conclusio
n
Against th
e
prod
uct
-
ba
se
d an
d u
s
e
r-b
ase
d
p
r
eju
d
i
c
e
s
in th
e traditional
coll
aborative
filtering algo
rithm and the
characte
ri
stic of ti
me-dat
a validity, this pap
er pro
posed a Tim
e
Weig
hted
Un
certai
n Neig
hborhoo
d Co
llaboration
Fi
ltering Algo
ri
thm (T
WUNCF). T
W
UNCF
effectively sol
v
es the p
r
obl
em of time-d
ata validity
with logis
t
ic
time func
tion. On this
bas
i
s, it
prop
oses an
idea
of un
certain
neig
h
b
o
rs,
and
dyn
a
mically
sel
e
cts th
e tru
s
t
w
orthy
neig
h
bors
with both a
u
s
er-b
ased m
e
thod an
d a
prod
uct
-
ba
se
d method, av
oiding the
un
certai
n a
c
curacy
of recomme
ndation u
n
d
e
r differe
nt occasi
on
s cau
s
ed by
simply usi
ng pro
d
u
c
t-based
recomme
ndat
ion o
r
use
r
-b
ase
d
recom
m
endatio
n. B
o
th the
expe
rimental
an
d
the the
o
retical
analysi
s
h
a
ve proved th
e
prop
osed
alg
o
rithm T
W
UNCF
ha
s
better a
c
cu
ra
cy and
stability tha
n
the traditional
algorithm TB
CF and IBCF.
Referen
ces
[1]
Jele
nc, Davi
d. Decisi
on mak
i
ng matters:
A
better
w
a
y to
evalu
a
te trust models.
Kn
o
w
ledge-B
a
se
d
System
s
. 20
13
; 52: 147-1
6
4
[2]
Che
n
MC, Ch
en LS, Hsu F
H
, et al. HPRS: A profitabili
t
y
bas
ed rec
o
mmend
er s
y
st
em. Martin
Hela
nd
er, Min
Xi
e, Rog
e
r Jia
o
, et al. Proce
edi
ng of
the IE
EE Internatio
n
a
l Co
nfere
n
ce
on Ind
u
strial
Engi
neer
in
g a
nd En
gin
eer
in
g Man
agem
en
t. Singap
ore,
IEEE Engin
e
e
r
ing Ma
na
gem
ent Soci
et
y
Sing
apor
e Ch
a
p
ter. 2007(
12):
219-2
2
3
[3]
Abeer
Moh
a
m
ed El-k
ora
n
y
,
Salma M
o
khta
r Khata
b
. Onto
log
y
-
base
d
S
o
cial
Recomm
e
nder
S
y
stem.
IAES International Jour
nal of Artificial In
telligence (IJ-AI).
2012; 1(3): 1
27-
138.
[4]
Park ST
, Pennock DM.
Apply
i
ng co
ll
abor
ativ
e filterin
g tech
niq
ues to
mov
i
e searc
h
for b
e
tter ranki
n
g
and br
ow
sing.
Proceed
in
gs of the 13th A
C
M SIGKDD In
ternatio
na
l Confer
ence
o
n
Kno
w
l
e
dg
e
Discover
y
and
Data Min
i
ng. K
DD 20
07, San
Jose, Ca
l
i
forni
a
, USA. ACM,
Ne
w
York 2
0
0
7
: 834–
84
3.
[5]
T
o
moharu I,
Kazumi
S, T
a
keshi
Y.
Modeling user behavior
in
re
c
o
mm
ender syst
em based on
max
i
mu
m e
n
t
r
opy
.
Procee
di
n
g
s of the 16th
Internation
a
l
C
onfer
ence
on
W
o
rld W
i
de W
eb. Nige
l
T
.
Banff, Alberta, Canada, World Scientific
Publishing Co.Pte
.Ltd. 2007: 1281-
1282
[6] Papp
ag
eorg
e
,
Charl
o
tte.
Rec
o
mmen
ded ve
rificatio
n
appr
o
a
ch for hu
ma
n
rated systems based o
n
lessons learned on the
orion launc
h abor
t system
. 2
1
st Annu
al Int
e
rn
ation
a
l S
y
m
p
o
s
ium of th
e
Internatio
na
l C
ounc
il on S
y
ste
m
s
Engin
eeri
n
g. 2011; 1: 1-1
3
.
[7]
Kakla
u
skas,
A. Recomme
nde
d b
i
ometri
c stress man
agem
ent s
y
st
em.
Expert System
s
wit
h
Appl
icatio
ns
. 2
011; 38(
11): 14
011-
140
25.
[8]
Massa P, Ave
s
ani P. T
r
ust-aw
a
r
e co
lla
bor
a
t
ive filterin
g for
recommen
der
s
y
stems
. Lect
u
re Notes
in
Co
mp
uter Scie
nce.
200
4; 329
0: 492-5
08.
[9]
Vincent SZ,
Boi Faltings.
Using
hi
erarc
h
ical c
l
usteri
n
g
for lear
ni
n
g
the o
n
tolo
g
i
es use
d
i
n
reco
mme
ndati
on systems.
Procee
din
g
s o
f
the 13th ACM S
IGKDD
Internatio
na
l Confer
ence o
n
Kno
w
l
e
d
ge D
i
s
c
over
y
an
d Dat
a
Mini
ng. San
Jo
se, Cal
i
forni
a
, United State
s
. 2007: 59
9-6
08.
[10]
Muhamm
ad W
a
seem
Ch
ug
htai, Al
i Bi
n S
e
l
a
mat, Imr
an Gh
ani. Go
al-
base
d
h
y
b
r
id
filter
in
g for
user-to-
user P
e
rson
al
i
z
ed
Recomm
e
ndati
on.
Inter
n
ation
a
l J
our
nal
of El
ectrical
and
Co
mputer
Eng
i
n
eeri
n
g
(IJECE).
2013;
3(3): 329-3
36.
[11]
GU Yi-ran,
C
H
EN Mi
n. On
e T
ag T
i
me-w
e
i
g
h
ted
Rec
o
mmen
d
Ap
pr
oach
on
T
r
ipartite Grap
hs
Net
w
orks.
Co
mp
uter Scie
nc
e.
2012; 3
9
(8): 96-9
8
,12
9
.
[12] HUANG
Ch
ua
ng-Gua
ng,
YIN Jian, W
A
N
G
Jing, et al.
Uncerta
i
n N
e
ig
hbors
’
Co
lla
bo
rative F
ilteri
n
g
Recomme
nd
ati
on Alg
o
rithm.
Chin
ese Jo
urn
a
l of Co
mp
uter
s.
2010; 33(8):
1369-
13
77.
[13]
CHEN Ji
an, YIN Jian. A C
o
ll
a
borativ
e F
ilteri
ng Rec
o
mme
n
datio
n Alg
o
rith
m Based o
n
In
fluenc
e Sets.
Journ
a
l of Softw
a
r
e. 20
07; 18
(7): 1685-
16
94
.
[14]
Liu
X, D
a
tta
A, Rzadca K,
Lim EP.
Ste
r
eo T
r
ust: A Group b
a
se
d
per-so
nal
i
z
e
d
trust mo
del.
Procee
din
g
s
of the
18th
ACM
c
onfer
enc
e o
n
In-formatio
n
a
nd K
n
o
w
l
e
dg
e
Mana
geme
n
t.
Hon
g
K
ong,
Chin
a. 20
09: 7
-
16.
[15]
Jamali M, Est
e
r M,
T
r
ust Walker.
A ran
d
o
m
w
a
lk
mo
del
for combin
in
g
trust-based
a
nd ite
m
-b
ase
d
reco
mme
ndati
on.
Proc
eed
in
gs of th
e 1
5
th
ACM SIGKD
D
Intern
ation
a
l
Conf
erenc
e o
n
Kn
o
w
l
e
d
g
e
Discover
y
and
Data Min
i
ng. P
a
ris, F
r
ance. 2
009: 39
7-4
06.
[16]
LOU Jun-g
a
n
g
,
SHEN Qing,
SHEN Z
han
g
-
guo. Ke
r
nel
Methods
of Soft
w
a
r
e
Re
li
abi
lit
y
Pr
edicti
on.
Co
mp
uter Scie
nce
. 201
2; 39(
4): 145-1
48.
[17]
Ma H, Ki
ng
I, L
y
u MR.
Effective
miss
in
g d
a
ta pr
edicti
on f
o
r
coll
ab
orative
fi
lterin
g.
Proce
e
din
g
s of th
e
30th Annual International ACM
SIGIR Conference. Amsterdam,
the Netherland
s. 2007: 39-46.
Evaluation Warning : The document was created with Spire.PDF for Python.