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.
9
, No
.
5
,
Octo
ber
201
9
, pp.
3813
~
38
21
IS
S
N:
20
88
-
8708
,
DOI: 10
.11
591/
ijece
.
v
9
i
5
.
pp3813
-
38
21
3813
Journ
al
h
om
e
page
:
http:
//
ia
es
core
.c
om/
journa
ls
/i
ndex.
ph
p/IJECE
Recomm
ender
S
ystem f
or
S
urplu
s
S
toc
k
C
l
earanc
e
Vipul
Agarwa
l, Vij
ayal
ak
sh
mi
A
Depa
rtment
o
f
C
om
pute
r
Scie
n
ce,
CHRIS
T
(De
e
m
ed
to
b
e
Unive
rsit
y
)
,
Ind
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
Ja
n 1
2,
2019
Re
vised A
pr 9,
2019
Accepte
d Apr
22
, 201
9
Acc
um
ula
ti
on
of
the
sto
ck
had
b
ee
n
a
m
aj
or
con
ce
rn
for
retail
shop
owners.
Surplus
stock
co
uld
be
m
ini
m
ize
d
if
th
e
s
y
s
te
m
coul
d
con
ti
nuou
sl
y
m
onit
or
the
accum
ula
t
e
d
stock
and
r
e
comm
end
th
ose
which
r
equi
re
c
le
ar
ance.
Rec
om
m
ende
r
S
y
stems
comput
e
s
the
da
ta,
shad
owing
the
m
anu
al
work
and
give
eff
i
ci
en
t
r
e
comm
enda
ti
ons
to
over
come
sto
ck
a
cc
um
ulation
,
cre
a
ti
ng
spac
e
for
n
ew
st
ock
for
sal
e
to
e
nhanc
e
th
e
pro
fi
t
in
business.
A
n
int
el
l
ige
n
t
rec
om
m
ende
r
s
ystem
was
bui
lt
th
at
coul
d
work
wi
th
th
e
data
and
h
el
p
the
shop
owners
to
over
c
om
e
the
issue
of
sur
plus
stock
in
a
remarka
b
le
w
a
y
.
A
n
item
-
it
em c
ol
la
bor
at
iv
e
filter
ing
t
ec
hni
que
with
Pe
arson
sim
il
ari
t
y
m
e
tr
ic
was use
d
to
dr
aw
th
e
sim
i
la
r
ity
be
twee
n
the
it
e
m
s
and
a
cc
or
dingly
g
iv
e
rec
om
m
enda
ti
on
s.
Th
e
result
s
ob
ta
in
ed
on
th
e
da
ta
set
high
li
ght
ed
th
e
top
-
N
it
ems
using
the
Pear
son
sim
il
arit
y
and
the
Cosin
e
sim
il
ari
t
y
.
Th
e
ite
m
s
havi
ng
the
highe
st
r
ank
had
th
e
highe
st
a
cc
um
ulation
a
nd
req
u
ir
ed
a
tt
e
nti
on
to
b
e
cl
e
are
d
.
Th
e
co
m
par
ison
is
dra
wn
for
the
pr
ec
is
ion
and
r
ec
a
ll
ob
ta
in
ed
b
y
th
e
sim
il
ari
t
y
m
e
tri
c
s
used.
Th
e
ev
aluati
on
of
th
e
ex
i
sting
work
was
done
using
pre
ci
sion
and
re
ca
l
l,
wher
e
th
e
p
rec
ision
ob
ta
in
e
d
was
remarka
bl
e,
whil
e
th
e
rec
a
ll
has
the
sc
ope
of
in
cre
m
en
t
but
in
turn
,
it
would
red
u
ce
th
e
v
al
ue
of
pre
ci
sion
.
Thus,
the
re
lies
a
scope
of
red
u
ci
ng
the
stock
a
cc
um
ulati
on
with
the
hel
p
of
a
re
comm
ende
r
s
y
st
em and ove
rco
m
e
los
ses t
o
m
axi
m
ize profit.
Ke
yw
or
d
s
:
Inve
nto
ry cle
a
r
ance
Re
com
m
end
er
syst
e
m
s
Re
ta
il
stock
m
anag
em
ent
Stock acc
um
ulati
on
Stock cl
eara
nc
e
Su
r
pl
us
st
ock
Copyright
©
201
9
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
:
Vipul A
ga
rw
al
,
Dep
a
rtm
ent o
f C
om
pu
te
r
Scie
nce,
CHRIST
(Dee
m
ed
to
be Un
i
ver
sit
y),
Ho
s
ur Mai
n R
oad, Be
ng
al
uru
-
56
0029, Kar
na
ta
ka,
India
.
Em
a
il
:
vip
ul.a
garwal@m
ca.chr
ist
unive
rsity
.in
1.
INTROD
U
CTION
Re
com
m
end
er
Syst
e
m
s
(RS)
hav
e
play
ed
a
vital
ro
le
to
s
peed
up
sim
pl
e
day
to
day
a
ct
ivit
ie
s
by
pro
vid
in
g
ef
fic
ie
nt
and
us
e
f
ul
reco
m
m
end
at
ion
s
be
it
in
the
e
-
com
m
erce
do
m
ai
n,
down
l
oa
ding
s
ongs
/
m
ov
ie
s
,
et
c
.
an
d
ha
ve
ne
ver
disap
poin
te
d
us.
D
ue
to
t
heir
ver
sat
il
it
y,
they
ha
ve
bec
om
e
an
i
m
po
rtant
top
ic
of
res
earc
h
and
i
ncr
easi
ngly
,
sta
rted
gr
a
bb
i
ng
t
he
us
e
r
’s
interest
.
St
ock
Ma
nag
em
en
t
is
the
basi
c
underst
an
ding
of
the
stoc
k
m
ix
in
a
c
om
pan
y
al
ong
with
var
i
ous
dem
and
s
of
the
stoc
k.
T
he
r
e
are
var
io
us
fa
ct
o
rs
w
hich
in
f
luence
the
st
ock
outfl
ow
a
nd
in
flo
w
that
in
tu
rn
af
fe
ct
the
pro
fit
an
d
l
os
s
in
t
he
or
gan
iz
at
io
n.
Alt
hough
the
re
ha
ve
bee
n
var
i
ou
s
m
et
ho
ds
incl
ud
i
ng
both
t
he m
anu
al
wor
k an
d sy
st
e
m
cal
culat
ion
,
ye
t
the
outc
om
e
is
so
le
ly
decided
by
the d
e
sig
nated
per
s
on.
Re
com
m
end
er
syst
e
m
s
ben
e
fit
both
t
he
us
e
r
and
the
pr
ov
i
de
r
by
sa
ving
t
he
tim
e
of
the
us
er
w
ho
keeps
on
searc
hing
for
an
it
e
m
wh
ic
h
they
prefe
r
in
t
he
a
pp
li
cat
ion
li
ke
e
-
c
omm
erc
e
w
hile
the
pr
ovide
rs
are
be
ne
fitt
ed
by enhanci
ng
t
heir
busi
ness
a
nd
p
opularit
y, ov
e
rall
incr
easi
ng
t
heir
ef
fici
ency.
The
st
oc
k
m
anag
em
ent syst
e
m
has
been
one of
the
c
r
ucial
prob
le
m
s
wh
ic
h
the b
us
i
ness
m
us
t
fo
c
us
on
to
m
axi
m
iz
e
their
pro
fit
an
d
m
ini
m
iz
e
loss.
W
it
h
te
ch
no
l
og
y
com
ing
to
pictu
re,
the
bu
si
ness
hu
bs
hav
e
m
igrated
from
the
pa
pe
r
work
to
c
om
pu
te
rs
in
the
form
of
e
xcel
s
heets
bu
t
sti
ll
no
t
m
a
king
e
ff
ic
ie
nt
us
e
of
t
he
act
ual
data,
fail
ing
in
data
a
na
ly
sis.
Larg
e
busi
ness
hubs
hav
e
a
hu
ge
am
ount
of
da
ta
in
e
xcel
s
he
et
bu
t
can
not
id
entify
the
surpl
us
st
ock
in
ad
va
nce
or
the
tre
nd
un
ti
l
the
ye
arly
,
quarterly
or
the
m
on
thly
analy
sis
is
done
,
an
d
by
the
ti
m
e
the
re
su
lt
s
are
out,
it
is
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.
9
, N
o.
5
,
Oct
ober
201
9
:
3813
-
3821
3814
too l
at
e to
reac
t. Mo
reove
r, t
he
se
kinds
of
work
r
e
qu
i
re
ad
di
ti
on
al
m
anp
ower
to
s
olv
e
th
e issue
a
nd t
hus ca
use
add
it
io
nal e
xp
e
ns
e t
o
the
busi
n
ess.
Hen
ce
,
r
ec
omm
end
er
syst
e
m
s
fo
r
s
urplu
s
stoc
k
cl
eara
nce
are
al
l
a
bout
m
anag
in
g
t
he
stoc
k
accum
ulati
on
,
wh
ic
h
will
stu
dy
an
d
analy
z
e
the
data
re
ga
rd
i
ng
t
he
sto
ck
in
th
e
in
ve
ntory
an
d
te
ll
about
the
top
-
N
it
em
s
bein
g
accum
ulate
d
in
the
organ
iz
at
io
n.
Alth
ou
gh
e
xcel
she
et
s
are
pr
e
sent
with
sim
il
ar
data,
ye
t
they
ha
ve
a
li
m
it
to
the
am
ou
nt
of
data
they
c
an
proce
ss
,
a
nd
they
al
s
o
la
c
k
t
he
m
ajor
a
dv
a
nc
e
m
ents/
al
gorithm
s
wh
ic
h
ca
n
be
ob
ta
ine
d
t
hrough
m
achine
le
arn
i
ng
al
gorith
m
s
than
the
ones
in
e
xcel
f
or
m
ula
co
m
pu
ta
ti
on
s
.
Th
us
, o
ne
m
us
t k
now
t
he use
of r
ec
omm
ender syst
em
in
this scena
rio
.
In
t
his
pa
per,
t
he
use
of
recom
m
end
at
ion
s
yst
e
m
is
being
pro
po
se
d
for
stock
cl
ea
ran
c
e,
w
hich
will
com
pu
te
the
to
p
-
N
it
em
s
wh
i
ch
ha
ve
a
ccum
ulate
d
i
n
the
stock
a
nd
needs
to
be
cl
ea
red
to
m
ini
m
iz
e
the
losse
s
du
e
to
s
urplu
s stock.
A
pyth
on
sc
ript
ha
s b
e
en
wr
it
te
n
wh
i
ch
trai
ns
the r
e
com
m
end
er
syst
e
m
with
the
e
xisti
ng
dataset
,
an
d
th
en
the
trai
ne
d
syst
e
m
is
fed
with
the
ne
w
da
ta
set
fo
r
inc
re
asi
ng
the
e
ff
ic
i
ency
of
t
he
pre
dicti
on
m
ade
by
the
a
lgorit
hm
.
The
al
gorithm
us
es
Pears
on’s
sim
i
la
rity
as
the
m
et
ric
fo
r
t
he
it
em
-
i
tem
colla
borati
ve
filt
ering
te
ch
niq
ue
to
draw
the
sim
il
arit
y
bet
ween
sim
i
la
r
ki
nd
of
it
em
s.
Once
the
data
is
fe
d,
t
he
c
orres
pondi
ng
resu
lt
s
a
re
obt
ai
ned
as
t
he
t
op
-
N
it
em
s
wh
i
ch
t
he
us
e
r
m
us
t
fo
c
us
to
cl
ear
the
st
ock
to
pr
e
ve
nt
lo
ss
es
an
d
m
axi
m
iz
e p
rofit
.
Along
with
t
hi
s,
the
syst
em
i
s
al
so
e
valuate
d
with
preci
sio
n
a
nd
recall
,
unde
rstan
ding
how
well
the
syst
e
m
is
pr
e
dicti
ng
a
nd
acc
ordi
ng
ly
dr
a
w
t
he
c
oncl
us
io
ns
f
or
the
syst
em
buil
t.
S
uch
kin
d
of
syst
em
do
es
n’
t
exist;
th
us
,
it
i
s
a
no
vel
idea
bein
g
pro
pose
d
t
o
e
ase
t
he
work
of
the
st
ock
m
anag
er
s
and
m
axi
m
iz
e
prof
it
by
ov
e
rc
om
ing
th
e losses
due t
o st
ock accum
ulati
on
.
Be
fore
c
reati
ng
a
reco
m
m
end
er
syst
em
,
one
m
us
t
be
fam
il
ia
r
with
th
e
basic
knowle
dge
a
bout
a
reco
m
m
end
er
s
yst
e
m
and
how
it
fu
nctio
ns
s
o
that
wh
il
e
re
plica
ti
ng
the
sam
e,
the
rese
arc
he
r
un
der
sta
nd
s
wh
at
,
wh
e
n
an
d
w
hy
ab
out
the
rec
omm
end
er
sys
tem
.
The
rec
om
m
end
at
ion
e
ng
i
ne
w
orks
ef
fici
ently
if
t
he
phases
pro
vid
e
a
su
cc
essive
ou
t
pu
t
f
or
the
ne
xt
ph
ase.
I
f
t
he
ou
t
pu
t
of
a
ny
of
t
hese
phases
is
com
pr
om
ise
d,
then
i
t
directl
y affects
the
qu
al
it
y o
f
t
he recom
m
end
at
ion
s.
Th
e
re a
re three
ph
a
ses
of the
r
ec
omm
end
at
io
n p
ro
ce
ss:
a.
Inform
at
ion
Colle
ct
ion
P
hase
This
phase
in
volves
the
c
ollec
ti
on
of
relat
ed
or
releva
nt
i
nfor
m
at
ion
[1
]
ab
out
t
he
us
e
r
pro
file
s
or
m
od
el
s
or
it
e
m
s
fo
r
the
pr
e
dicti
on
ta
sk
w
hich
in
volves
us
er
’s
at
trib
utes,
be
hav
i
or
s
,
li
kes,
disli
kes,
e
xp
li
ci
t
feedbac
ks
,
et
c
.
b.
Learn
i
ng P
has
e
This
phase
re
quires
a
le
ar
ning
al
gorithm
to
filt
er
an
d
e
xp
l
oit
the
us
er
’s
f
eat
ur
es
f
r
om
t
he
fee
db
a
c
k
gathe
red
in
t
he
inf
or
m
at
ion
co
ll
ect
ion
phase
a
nd
in
tur
n
fin
d
ou
t
a
bout
the
re
le
van
ce
in
t
he
e
xisti
ng
in
form
a
ti
on.
c.
Pr
e
dicti
on
or R
ecom
m
end
at
io
n
P
hase
This
ph
ase
is
a
lso
kn
own
as
the
im
ple
m
enta
ti
on
ph
ase
w
hi
ch
rec
omm
end
s
or
pre
dicts
w
hat
ki
nd
of
it
e
m
s
the
us
e
r
m
ay
pr
efe
r,
de
pendin
g
on
the
inf
orm
ation
c
ollec
te
d
i
n
th
e
first
phase
a
nd
analy
sis
in
vo
l
ve
d
(b
y
us
in
g
t
he
al
gor
it
h
m
s,
existi
ng
on
e
s
or
t
he ne
w on
es
)
i
n
the
seco
nd p
hase.
2.
RECOM
ME
NDATIO
N
F
I
LT
ERING T
ECHNIQ
UES
The
rec
omm
end
at
ion
syst
em
so
le
ly
de
pends
on
t
he
diff
e
re
nt
filt
erin
g
te
c
hn
i
qu
e
s
us
ed
t
o
get
ef
fici
ent
reco
m
m
end
at
ion
s
.
The
re
a
re
seve
ral
te
c
hn
i
qu
e
s
a
vaila
ble
an
d
e
ach
te
ch
nique
se
r
ves
di
ff
ere
nt
pur
po
s
es
an
d
highli
gh
ts
their
potenti
al
in va
rio
us
do
m
ai
ns
dep
e
ndin
g on the
requirem
ent.
2.1.
Conten
t
-
base
d f
il
terin
g
In
this
filt
erin
g
te
ch
nique,
a
m
at
ch i
s
m
ade
be
tween
the
de
s
cripti
on
of
it
em
s
and
descr
i
pt
ion
giv
e
n
by
us
ers
f
or
the
it
e
m
wh
ic
h
they
are
sea
rc
hing.
The
rec
omm
e
nd
at
io
ns
so
le
ly
li
e
on
w
hat
t
he
cu
rr
e
nt
use
r
seeks
dep
e
ndin
g
on
it
s
descr
i
ptio
n
r
at
her
t
han
gi
vin
g
a
rec
omm
e
nd
at
io
n
base
d
on
the
oth
e
r
use
rs
of
t
he
sam
e
kind
,
li
ke
that
in
c
ollaborat
ive
te
ch
nique.
It
invol
ves
our
work
on
t
he
m
et
adat
a.
This
process
do
e
s
not
re
quire
t
he
prof
il
e
of
oth
e
r
us
ers
as
it
does
not
in
flue
nce
reco
m
m
end
at
ions.
If
the
cu
rr
e
nt
us
er
s’
pro
file
ch
an
ge
s,
th
e
te
chn
iq
ue
has
t
he
ca
pab
il
it
y
to
a
dju
st
it
s
rec
omm
end
at
io
ns
in
neg
li
gi
ble
tim
e.
The
m
ai
n
disad
va
ntage
of
t
his
te
chn
iq
ue
is
th
at
it
m
us
t
hav
e
in
-
de
pth
kn
owle
dge
a
bout
t
he
desc
riptio
n
of
al
l
the
featu
res
i
n
the
prof
i
le
el
se
the p
e
rfo
rm
ance goes
dow
n
si
gn
i
ficantl
y.
2.2.
Co
ll
abor
ati
ve f
il
tering (
CF)
This
te
ch
nique
giv
es
rec
omm
end
at
io
ns
base
d
upon
the
sim
il
ariti
es
betw
een
the
us
e
r
a
nd
bet
wee
n
the
it
em
s
bein
g
s
earc
hed
de
pendin
g
on
ex
plici
t
releva
nc
e
(
rati
ngs,
ta
gs,
et
c.
)
or
im
pli
ci
t
releva
nce
(
act
ion
s
involve
d
li
ke
rea
ding,
do
wn
l
oad
i
ng,
et
c.
).
It
de
pends
on
t
he
prefere
nces
of
oth
e
r
us
er
s
t
o
dra
w
reco
m
m
end
at
ion
s
.
A
us
er
wi
ll
get
a
reco
m
m
end
at
ion
of
t
ho
s
e
it
em
s
that
ha
ve
no
t
be
en
rate
d
be
fore
but
wer
e
rated
by
oth
e
r
us
ers
i
n
the
ne
ighbor
hood
i.e.
tho
se
us
er
s
w
ho
ha
ve
sim
il
a
r
interest
s
a
nd
pr
e
fer
e
nces
whic
h
is
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
Reco
mm
e
nder
syste
m
fo
r s
ur
pl
us
stock
clear
an
ce
(
Vipu
l
Ag
ar
wal
)
3815
cal
culat
ed
by
the
sim
il
arities
betwee
n
thei
r
pro
file
s.
As
per
Figure
1,
we
ca
n
see
the
bas
ic
unde
rstan
ding
of
the
CF an
d
c
onte
nt
-
base
d fil
te
rin
g. T
his tech
niqu
e is o
ne of
the
m
os
t widely
im
ple
m
ented.
Figure
1. Coll
aborati
ve fil
te
ring and c
onte
nt
-
base
d fil
te
rin
g
3.
RELATE
D
W
ORK
Re
com
m
end
er
syst
e
m
has
bee
n
a
m
ajo
r
t
op
i
c
of
resea
rch
over
t
he
past
de
cades.
Va
rio
us
al
gorithm
s
hav
e
bee
n
de
pl
oyed
e
xtensi
ve
ly
in
var
io
us
dom
ai
ns
to
enh
a
nce th
e
perfor
m
ance o
f
the
r
ecom
m
end
er sy
stem
s.
The
hum
an
-
rec
omm
end
er
inte
racti
on
s
base
d
on
visu
al
iz
at
i
on
fr
am
ework
[2]
w
hich
com
bin
es
with
visu
al
iz
at
ion
te
chn
iq
ue
wa
s
pro
po
se
d.
Depend
i
ng
on
f
urt
her
analy
sis
of
the
existi
ng
s
yst
e
m
s
and
res
ult
sur
veys,
t
he
pa
pe
r
dr
a
ws
f
uture
r
esearch
c
halle
ng
e
s
an
d
op
portu
niti
es
in
this
area.
A
surve
y
on
e
-
c
omm
e
rce
sit
es
[3
]
us
ing
R
S
was
perf
or
m
ed
to
know
how
they
are
bei
ng
ben
e
fite
d
by
i
nc
reasin
g
t
heir
sal
es
an
d
e
nha
ncin
g
t
heir
pro
fit.
It
al
so
desc
ribes
var
i
ou
s
sit
es
wh
ic
h
ha
ve
in
corp
or
at
ed
m
or
e
tha
n
one
re
com
m
end
er
sy
stem
.
A
br
ie
f
stud
y
i
s
m
ade
on
the
inputs
require
d
by
the
m
from
the
c
us
tom
ers,
the
reco
m
m
e
nd
at
io
n
te
c
hn
i
qu
e
s
an
d
t
he
i
nterf
ac
e
pro
vid
e
d
to
t
he
us
e
rs.
E
-
c
omm
erce
has
al
lo
wed
pe
ople
to
choose
from
m
ulti
ple
opti
ons,
br
ea
king
t
he
st
ereo
ty
pe
of m
ass p
rod
uc
ti
on
.
Web
us
age
cl
ust
ering
has
bee
n
a
po
pu
la
r
ar
e
a
of
researc
h
w
her
ei
n
diff
e
re
nt
te
chn
i
qu
es
ar
e
bei
ng
us
e
d
to
get
the
m
axi
m
u
m
ou
tp
ut
from
web
us
a
ge
[4
]
.
T
he
f
ocus
li
es
on
ov
e
rco
m
ing
the
obsta
cl
es
,
giv
i
ng
t
he
c
orrect
resu
lt
of
we
b
us
a
ge
cl
us
te
r
ing
usi
ng
rec
omm
end
at
io
n
syst
e
m
s.
The
process
of
fi
ndin
g,
analy
zi
ng,
pre
-
p
r
ocessi
ng
an
d
e
ff
ic
ie
nt
cl
us
te
rin
g
of
da
ta
helps
in
the
bu
il
ding
eff
ic
ie
nt
rec
omm
end
er
syst
e
m
s.
The
a
utho
r
pro
po
s
es
a
n
a
naly
sis
an
d
pr
e
-
pro
cessi
ng
m
od
el
to
c
om
e
ou
t
of
the
poor
qu
al
it
y
of
we
b
us
a
ge
data,
wh
e
reas
to
e
nsure
th
e
e
ff
ic
i
ency
of
th
e
cl
us
te
rin
g,
Partic
le
Swa
rm
Optim
iz
at
ion
(P
S
O)
a
ppro
a
ch
i
s
us
ed
.
Com
m
on
ly
kn
own
as
KDD
(
Kno
wled
ge
Di
sco
ver
y
in
Data
m
ining
)
,
the
te
rm
ho
lds
gre
at
i
m
po
rtance
in
the
extracti
on
of
use
fu
l
an
d
rele
va
nt
in
f
or
m
at
ion
form
a
la
rg
e
da
ta
set
.
T
hese
w
ork
on
the
basi
s
of
cl
us
te
ri
n
g
r
el
at
ed
data,
patte
r
ns
,
pr
e
dicti
on
of
requirem
ents
of
us
er
s
an
d
giv
in
g
valid
inf
or
m
at
ion
.
T
hu
s
,
a
stu
dy
of
thei
r
requirem
ent,
th
e
patte
r
ns
f
ollo
wed
by
them
and
thei
r
i
nteres
ts
is
esse
ntial
f
or
we
b
usa
ge
c
lusterin
g.
disc
usse
s
the d
i
ff
e
ren
t i
m
age classi
fic
at
ion
alg
ori
thm
s
.
The
im
po
rtanc
e
of
cl
us
te
ri
ng
of
we
b
sessi
ons
[
5]
is
draw
n,
wh
ic
h
is
a
n
integral
par
t
of
the
m
ining
te
chn
iq
ue
to
group t
he se
s
sio
ns
base
d
on
s
ome
sim
il
arit
y
bet
ween
them
. T
he
pa
per
al
so
gi
ves a
ne
w al
gorithm
for
Pa
rtic
le
S
war
m
Op
ti
m
izati
on
(P
S
O).
T
he
pro
po
se
d
al
gorithm
has
no
li
nk
with
ot
her
existi
ng
cl
us
te
rin
g
al
gorithm
s.
The
m
ai
n
j
ob
of
web
sessio
n
cl
us
te
rin
g
is
to
e
xtract
the
t
otal
us
a
ge
of
the
web
a
nd
the
pa
tt
ern
in
wh
ic
h
the
user
nav
i
gates
betw
een
web p
a
ges
and pre
dict t
he
b
e
hav
i
or.
A
n
Ultra
Lar
ge
Scal
e
(U
L
S)
so
ft
war
e
pro
je
ct
s
[6
]
are
c
onside
red
t
o
ha
ve
hi
gh
c
om
plexity
,
ye
t
the
requirem
ents
and
nee
ds
of
the
se
pro
j
ect
s
a
re
no
t
m
et
.
Th
us
,
a
proces
s
is
being
est
ablishe
d
t
o
us
e
th
e
da
ta
m
inin
g
and
re
com
m
end
er
syst
em
s
to
involve
the
sta
keho
lde
rs
of
s
uch
la
r
ge
pro
j
e
ct
s.
Data
Mi
ni
ng
is
c
onside
re
d
as
a
us
ef
ul
te
c
hn
i
que
w
he
rein
the
da
ta
is
c
ollec
te
d,
a
nd
they
are
s
tructu
red
a
nd
c
at
egorized
f
or
easy
proce
ssin
g
a
nd
ou
tc
om
e.
Atte
ntion
is
draw
n
towa
rd
s
the
hig
hlig
hts
of
t
he
reco
m
m
end
at
ion
s
on
per
s
on
a
li
zed
pr
oducts
[7
]
a
nd
to cross
the
hur
dle of i
nfor
m
ation
overl
oad when
the am
ount
o
f
d
at
a i
nvolve
d
is
huge
.
Ov
e
r
the
ye
ars
,
E
-
com
m
erce
has
f
un
ct
io
ne
d
di
ff
e
ren
tl
y
a
nd
now
prov
i
de
s
rec
omm
end
at
ion
s
m
or
e
li
kely
and
ef
fic
ie
nt,
e
nh
a
ncin
g
their
busine
ss.
Re
com
m
end
er
s
yst
e
m
s
in
e
-
c
om
m
erce
[8
]
portrays
the
us
e
of
the
e
m
erg
in
g
syst
e
m
in
the
e
-
com
m
erce
industry,
on
e
of
the
fastest
gr
owin
g
m
ark
et
s
in
tod
ay
’s
world
.
To
co
m
pete
with th
e
gro
wing
virtu
al
m
ark
et
and th
e
physi
cal
m
ark
et
, th
e
s
yst
e
m
is ada
pted
i
n suc
h a
wa
y t
hat
tough
c
om
petition
is
ens
ure
d
betwee
n
t
he
tw
o
in
dustrie
s,
hav
i
ng
the
sam
e
go
al
s
bu
t
dif
fer
e
nt
a
ppro
ac
h.
Th
us
, it
dr
a
ws an in
sig
ht into
the tec
hniq
ues t
o
pr
ov
i
de
rec
om
m
end
at
ion
s t
o
the
cust
om
ers.
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.
9
, N
o.
5
,
Oct
ober
201
9
:
3813
-
3821
3816
A
m
ov
ie
-
base
d
r
ec
omm
end
at
i
on
s
yst
em
[9
]
i
s
a
bo
on
f
or
s
oc
ie
ty
.
The
k
-
m
ean
cl
us
te
r
is
use
d
in
this
case
w
her
ei
n
t
he
cl
us
te
r
hea
ds
a
re
m
ade
of
si
m
il
ar
kin
d
of
it
em
s
and
gr
oupe
d
to
gethe
r
.
Eac
h
obse
rv
a
ti
on
is
bro
ken
int
o
k
-
cl
us
te
rs
w
her
e
each
cl
us
te
r
belo
ngs
t
o
t
he
cl
us
te
r
ha
ving
t
h
e
nea
rest
m
ean,
wh
ic
h
s
at
isfie
s
the pr
op
e
rty
of
the clu
ste
r.
He
nce, usi
ng this,
an
att
em
pt is m
ade to
in
crea
se the e
ff
ic
ie
nc
y of t
he
syst
em
.
Su
c
h
a
pool
of
us
es
of
rec
omm
end
er
syst
e
m
s
has
bee
n
discuss
e
d
al
on
g
with
their
ad
va
ntages
a
nd
disad
va
ntages
in
their
res
pect
ive
fiel
d.
T
he
cold
-
sta
rt
iss
ue
as
we
know
is
one
of
the
m
ajo
r
draw
bac
ks
i
n
a
ny
reco
m
m
end
er
syst
e
m
s.
Her
e
we
cl
assi
fy
c
old
-
sta
rt
prob
le
m
into
two
cat
egories
,
i.e.
co
ld
-
us
er
a
nd
c
old
-
it
em
.
A
c
old
-
us
e
r
is
on
e
w
ho
is
a
ne
w
us
er
[
10
]
t
o
the
syst
em
.
Fo
r
insta
nce
,
wh
e
n
a
us
e
r
c
r
eat
es
a
ne
w
ac
count
i
n
any
of
the
e
-
c
omm
erce
sit
es
f
or
bu
yi
ng
an
it
e
m
,
the
syst
e
m
is
una
ble
t
o
i
nteract
with
the
use
r
to
prov
i
de
a
reco
m
m
end
at
ion
as
t
her
e
is
da
ta
inade
quacy
ab
ou
t
t
he
us
er
’s
histor
y,
li
ke
s,
disli
ke
s
a
nd p
re
vious pur
c
ha
ses
in
oth
e
r
sim
i
la
r
dom
ai
n
s.
Ther
e
is
an
at
te
m
pt
t
o
re
duce
the
s
par
sit
y
issue
al
so
i.e.
re
du
ci
ng
the
s
pa
rsity
of
t
he
sp
ars
e
m
at
rix.
This
w
ou
l
d
fill
in
the
ho
le
s
in
the
s
pa
rse
m
at
rix
s
o
that
t
her
e
is
no
data
ins
uff
ic
ie
ncy
an
d
t
he
col
d
sta
rt issue is
ov
erco
m
e.
4.
DA
T
AS
ET
O
VER
VIEW
A
da
ta
set
is
bu
il
t
fo
r
im
ple
m
e
nting
the
pro
pose
d
s
olu
ti
on
f
or
overc
om
ing
the
acc
um
ulatio
n
of
stoc
k.
The
da
ta
set
includes
t
he
f
ollow
i
ng
set
of
a
tt
ribu
te
s:
us
e
r
id,
it
e
m
id,
purch
ase
d,
st
oc
k
and
date.
T
he
dataset
include
s
the
it
em
-
id
of
le
t'
s
sa
y,
deterg
e
nts
of
dif
fer
e
nt
bran
ds
. T
he data
is
store
d i
n t
he
form
of
rows,
wherein
each
r
ow
has
entries
f
or
the
above
at
tri
bu
te
s,
i.e.
it
em
-
wise
the
detai
ls
ar
e
store
d,
a
nd
t
he
us
er
-
id
i
nd
i
cat
es
inf
or
m
at
ion
ab
ou
t
wh
ic
h
of
t
he
c
us
tom
er
ha
s
purc
hase
d
the
c
orrespo
nd
i
ng
it
em
al
on
g
with
the
date
a
nd
t
he
rem
ai
nin
g
stoc
k
of
that
it
em
.
The
dataset
ha
s
ent
ries
f
or
a
pe
rio
d
of
over
tw
o
ye
ars.
The
at
tri
bu
te
“
stock”
,
is
the
m
ai
n
at
trib
ute
of
t
he
da
ta
set
.
This
at
t
rib
ute
will
sto
r
e
the
c
ount
of
ever
y
pro
du
ct
wh
ic
h
is
le
ft
in
the
inv
e
ntory
. Hen
ce the syst
em
sh
al
l rely
on t
his m
ai
n
at
tribu
te
to pr
ovide a
n effici
ent
r
ec
om
m
end
at
ion
.
5.
PROP
OSE
D MET
HO
DOL
OGY
Ther
e
are
rec
omm
end
at
io
ns
syst
e
m
s
bu
il
t
for
e
-
com
m
e
rce,
m
ov
ie
,
et
c.
I
n
this
res
earch
w
ork
,
base
d
on
the
li
t
eratur
e
re
view
,
a
syst
em
is
design
e
d
t
hat
gi
ve
s
rec
omm
end
at
ion
s f
or
cl
ea
r
ing
t
he
s
urpl
us
stoc
k
in
the
in
ve
ntor
y
in
retai
l
s
hops
,
m
al
ls,
and
m
ega
-
m
arts.
T
he
syst
e
m
sh
al
l
rec
omm
end
t
op
-
N
it
em
s
hav
in
g
a
m
axi
m
u
m
accum
ulati
on
in
the
i
nv
e
ntory
.
As
data
is
generate
d
at
a
dras
ti
c
sp
ee
d,
rec
o
m
m
end
er
syst
e
m
s
hav
e
sign
ific
a
ntly
re
du
ce
d
t
he
m
anu
al
w
ork
a
nd
drast
ic
al
ly
de
m
on
st
rates
the r
e
qu
i
red
outp
ut
by
prov
i
ding
ef
f
ic
ie
nt
reco
m
m
end
at
ion
s
i
n
var
i
ous
do
m
ai
ns
li
ke
e
-
com
m
erce
sit
es
(
Am
azon
,
F
li
pk
art)
,
entert
ai
nm
ent
(N
et
fl
ix),
et
c
.
Figure
2 dem
on
str
at
es a
n
a
rc
hitec
ture dia
gr
a
m
f
or the
pro
po
s
ed
work.
Figure
2. The
pro
po
se
d
m
et
ho
d f
or the
Item
-
Item
si
m
il
arity
m
et
ric
The
a
bove
a
rc
hitec
ture be
gins wit
h
th
e d
at
a
set
w
hic
h
s
houl
d
be divi
ded into t
he
trai
ni
ng d
at
a an
d
t
he
te
sti
ng
data,
s
o
to
go
with
t
he
us
ua
l
te
chn
i
que
us
ed
i
n
m
achine
le
arn
i
ng,
th
e
first
half
or
quarter
of
the
da
ta
set
can
be
us
e
d
as
t
he
trai
ning
dat
a
an
d
th
e
rem
ain
in
g
ca
n
be
us
e
d
as
the
te
sti
ng
data.
Ba
se
d
on
this,
on
c
e
the
m
od
el
is
trai
ned,
we
pro
vid
e
the
m
od
el
i
nput
data
an
d
the
ne
w
input
data
as
th
e
te
sti
ng
data
to
the
t
raine
d
m
achine
le
arn
in
g
al
gori
thm
.
The
res
ul
ts
ob
ta
ine
d
ar
e
the
pr
e
dicti
ons
of
the
to
p
-
N
it
em
s
hav
in
g
the
highest
stock
accum
ulati
on
,
wh
ic
h
nee
ds
to
be
cl
eare
d,
ranked
from
the
hig
he
st
pr
io
rity
to
t
he
l
owest
pr
i
or
it
y
to
a
vo
i
d
l
os
se
s
.
Anothe
r
prospect
to
this
m
a
y
be
seen
if
w
e
con
si
der
f
or
a
longer
per
i
od.
I
f
these
it
e
m
s
hav
e
po
or
ou
t
flo
w
in
a
ye
ar
but
gr
a
du
al
l
y
i
m
pr
ov
e
s
in
t
he
c
on
se
quent
ye
ars,
the
rec
om
m
end
er
le
ar
ns
this,
a
nd
al
th
ough
a
par
ti
cula
r
it
e
m
m
ay
hav
e
le
sser
sal
e
com
pared
to
a
no
t
her
i
tem
,
the
sal
e
gr
ad
ually
picks
up
,
th
us
the
ra
nk
of
popula
rity
cha
ng
e
s
f
or
that
i
tem
desp
it
e
ha
ving
m
or
e
st
oc
k
as
in
f
utu
r
e,
an
d
the
sal
e
m
igh
t
inc
rea
se
f
or
that
it
e
m
.
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
Reco
mm
e
nder
syste
m
fo
r s
ur
pl
us
stock
clear
an
ce
(
Vipu
l
Ag
ar
wal
)
3817
Hen
ce
,
t
he
store
s
ca
n
c
om
e
up
with
their
own
set
of
m
echan
ism
to
over
com
e
stock
ac
cum
ulati
on
as
they
will
be
ge
tt
ing
a
pr
io
r
reco
m
m
end
at
ion
to
c
op
e
up
with
the
plen
ty
stock
a
vaila
ble
in
th
e
i
nvento
ry.
Hen
ce
f
or
eac
h
cat
eg
or
y
li
ke
biscuits,
det
erg
e
nts,
c
ho
c
ol
at
es,
j
uices
,
e
tc
.,
the
rec
omm
end
at
ion
en
gi
ne
can
pro
vid
e
gr
eat
r
esults t
o
cl
ea
r
t
he
s
urplu
s stoc
k.
It
is
seen
t
hat
the
ta
rg
et
at
tri
bu
te
“St
oc
k”
will
play
a
vital
ro
le
in
a
nal
yz
ing
the
e
xis
t
ing
st
ock
i
n
the
in
ve
ntory
.
The
it
e
m
-
it
e
m
si
m
il
arity
us
in
g
Pears
on’s
c
orrelat
ion
as
it
f
ind
s
the
relat
ion
sh
i
p
bet
ween
t
w
o
var
ia
bles,
i.e.,
in
o
the
r
w
ords
,
it
ind
ic
at
es
th
e
stren
gth
of
t
he
relat
io
ns
hi
p.
W
e
are
al
s
o
t
ryi
ng
th
e
sam
e
with
Cosine
sim
il
ari
ty
wh
ere
the
C
os
ine
is
the
a
ngle
between
a
ny
two
it
em
vecto
rs,
in
dicat
ing
t
hat
cl
os
er
t
he
ve
ct
or
,
la
rg
er
will
the
Cosine
a
nd
sm
al
le
r
the
a
ng
le
.
The
s
im
i
la
rity
will
be
draw
n
for
it
em
s
of
th
e
sam
e
cat
ego
r
y
and
thu
s
it
e
m
-
it
e
m
m
a
trix
will
be
create
d
by
the
al
go
rithm
t
o
process
f
ur
t
her.
T
h
is
m
at
ri
x
will
help
to
identify
the
sim
i
la
r
kind
of
it
e
m
s
an
d
fig
ur
e
out
w
hich
is
the
m
ax
im
u
m
stock
le
ft
i.e.
s
howi
ng
th
e
le
ast
out
flo
w.
Hen
ce
,
a
n
ef
fic
ie
nt
it
em
-
si
m
ilarity
m
e
tric
can
be
use
d
to
cr
eat
e
a
reco
m
m
end
at
io
n
e
ng
i
ne
f
or
cl
eari
ng
t
he
stoc
k
from
the
in
ve
nt
or
y
a
nd
overc
om
e
losses
a
nd
m
axi
m
izing
t
he
pr
of
it
.
A
syst
e
m
li
ke
this
does
no
t
exist,
he
nce
it
can
be
inc
orp
orat
ed
i
n
the
sal
es dom
ai
n
for
r
et
ai
l shops,
m
al
ls, an
d
m
ega
-
m
arts.
6.
IMPLEME
N
TATION
AN
D RESULTS
In
this
pa
per
,
the
it
em
-
it
e
m
c
ollaborat
ive
filt
ering
is
bein
g
us
e
d
wh
e
rein
a
n
it
em
loo
k
al
i
ke
m
at
rix
is
create
d,
w
hich
will
be
f
ur
the
r
processe
d,
us
i
ng
w
hich
r
eco
m
m
end
at
ion
c
an
be
giv
e
n.
H
ence,
with
the
si
m
il
ar
kind
of
it
em
s,
we
get
the
to
p
-
N
rec
omm
end
at
ion
of
t
hose
it
e
m
s
wh
ic
h
ha
ve
a
ccum
ulate
d
over
a
pe
ri
od
a
nd
accor
dingly
di
ff
e
ren
t
m
echan
ism
s
can
be
adopted
by
the
m
anag
em
e
nt
t
o
ta
ke
car
e
of
thei
r
inve
ntory
.
The
al
gorithm
create
s
a
sim
i
la
rity
m
at
rix
by
the
pair
of
it
e
m
s
that
ha
ve
stoc
k
of
it
em
s
dec
reasi
ng
by
th
e
purc
hase
s
m
ade
by
the
us
e
rs
to
ease
the
pro
cess
a
nd
m
ake
the
m
od
el
m
or
e
reli
able.
Th
e
pse
ud
ocode
of
the
script
us
e
d
is
gi
ven
belo
w:
Step
1: I
m
po
rt
necess
a
ry li
br
a
ries
Step
2: Read
th
e d
at
aset
f
il
e
Step
3: Pa
rtit
ion
the
d
at
aset
i
nt
o
a
2
:1
rati
o f
or trainin
g
a
nd
te
sti
ng
r
e
sp
ect
i
vely
.
Step
4: Tr
ai
n
t
he
m
od
el
w
it
h t
he
Trai
ning
da
ta
w
it
h
"t
ar
get
=purc
hased
"
a
nd "sim
i
la
rity
_type=Pears
on
"
.
Step
5: Tr
ai
n
t
he
m
od
el
wit
h t
he
Trai
ning
da
ta
w
it
h
"t
ar
get
=purc
hased
"
a
nd "sim
i
la
rity
_type=
Cosine
".
Step
6: Make
predict
io
ns
with
the Test
data a
nd print t
op
-
N
it
e
m
s.
Step
7: E
valuat
e preci
sion_
rec
al
l.
To
a
vo
i
d
com
m
on
m
ist
akes
in
re
plica
ti
ng
or
c
reati
ng
a
r
ecom
m
end
er
s
yst
e
m
,
fo
r
this
prob
le
m
in
sp
eci
fic
or
ot
he
rs
in
ge
ne
ral,
on
e
m
us
t
m
ake
su
re
that
the
tr
ai
nin
g
data
an
d
te
sti
ng
data
are
div
i
ded
i
n
at
le
ast
2:1
rati
o.
i.e.
a
highe
r
nu
m
ber
of
data
row
s
f
or
t
he
trai
ni
ng
dataset
a
nd
a
le
sser
num
ber
of
data
r
ows
for
the
te
sti
ng
dat
as
et
.
This
pro
vid
es
a
n
e
ff
ect
ive
trai
ning
of
the
rec
omm
e
nd
e
r
syst
em
a
nd
give
s
bette
r
resu
lt
s
irres
pecti
ve
of
the
siz
e
of
the
dataset
use
d
f
or
the
syst
em
in
the
f
uture
.
Als
o,
t
he
dataset
m
us
t
be p
re
-
processe
d
thoro
ughly
to
a
vo
i
d
m
issi
ng
va
lues,
outl
ie
rs,
or
a
ny
s
ort
of
noise
,
al
th
ough
no
ise
w
on’t
im
pact
the
syst
em
m
uch
sti
ll
it
is adv
ise
d
to
r
em
ov
e t
hem
from
the
da
ta
set
f
or
sm
oo
th
functi
onin
g.
The
scri
pt
cre
at
ed
is
a
bit
tim
e
con
su
m
ing,
on
ave
rag
e
a
rou
nd
2
m
inu
te
s
for
the
e
xisti
ng
dataset
,
and
it
m
a
y
vary
du
e
to
t
he
a
m
ou
nt
of
data
it
will
process
to
giv
e
a
rec
om
m
end
at
ion
.
A
syst
em
with
bette
r
hard
war
e
c
onfi
gurati
on,
wh
ic
h
m
os
t
of
the
pe
op
le
hav
e
i
n
tod
ay
’s
tim
e,
the
processi
ng
w
il
l
be
m
uch
fast
er
tha
n
the
on
e
do
ne
he
re
as
it
was
de
li
ber
at
el
y
w
orked
on
a
sta
nda
rd
syst
em
to
set
an
e
xam
ple
for
t
he
e
nd
us
e
rs
t
hat
m
ini
m
al
req
uir
e
m
ents
will
ser
ve
the
sam
e
pu
rpose
as
t
hose
with
the
la
te
st
t
echnolo
gy.
O
n
execu
ti
ng
t
he
s
cript,
the foll
owin
g o
utput i
s
ob
ta
in
ed:
Table
1
s
hows
that
f
r
om
the
va
rio
us
num
ber
of
it
em
s,
the
ra
n
k
po
i
nts
out
t
o
t
he
sta
tus
of
it
e
m
s
being
so
ld
,
i.e.
rank
1
in
dicat
es
the
it
e
m
-
id
103
ha
s
bee
n
sig
nific
antly
been
acc
um
ulate
d
in
th
e
inv
e
ntory
,
w
hile
the
it
e
m
-
id
101
wi
th
ra
nk
5,
has
go
t
the
le
ast
ac
cum
ulati
on
of
the
stoc
k
in
th
e
inv
e
ntory
.
N
ext,
we
t
ry
to
execu
t
e
the
sc
ript
with
the
Cosi
ne
Sim
il
arity
wh
e
re
Ta
ble
2
s
how
s
that
f
ro
m
the
va
rio
us
num
ber
of
it
em
s,
the
ra
nk
po
i
nts
out
to
t
he
sta
tus
of
it
em
s
being
so
l
d,
i.e.
ra
nk
1
i
ndic
at
es
the
it
e
m
-
id
103
has
be
en
sig
nifica
ntly
been
accum
ulate
d i
n t
he i
nv
e
nt
or
y,
wh
il
e
the
it
em
-
id
101
with
rank
5,
ha
s
got
the
le
ast
accum
ul
at
ion
of
the
st
oc
k i
n
the in
ven
t
or
y.
Howe
ver, there
is a f
li
p
in t
he
r
an
k of
it
em
-
id 102 an
d
it
em
-
id 1
04.
N
ow
let
u
s r
ef
er to
Figure
3
and g
et
a
n u
nderstan
ding
of
the
belo
w
-
obta
ined
tables.
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.
9
, N
o.
5
,
Oct
ober
201
9
:
3813
-
3821
3818
Table
1.
T
he
r
a
nk ass
ociat
ed wit
h
it
em
s
in stoc
k usin
g
Pears
on
'
s
sim
i
l
arit
y
Ite
m
-
id
Ran
k
103
1
104
2
102
3
105
4
101
5
Table
2.
T
he
r
a
nk ass
ociat
ed wit
h
it
em
s
in stoc
k usin
g
Cosine
sim
il
ari
ty
Ite
m
-
id
Ran
k
103
1
102
2
104
3
105
4
101
5
Figure
3. Yea
r wise i
te
m
sale
Com
par
ing
the
gr
a
ph
data
wi
th
the
res
ult
obta
ined
i
n
Ta
bl
e
1,
we
ca
n
s
ee
that
al
tho
ug
h
the
re
is
a
sign
ific
a
nt
sal
e
of
the
it
em
-
id
103
c
om
par
ed
with
it
em
-
id
104
an
d
10
2,
ye
t
it
has
rank
1.
This
is
be
caus
e
the
reco
m
m
end
er
s
yst
e
m
le
arn
s,
or
in
ot
her
words,
we
ca
n
say
that
the
al
gorith
m
s
le
arn
from
t
he
data
a
nd
th
us
they
can
ide
ntify
that
the
stock
ac
cum
ulati
on
is
i
ncr
easi
ng
for
it
e
m
-
id
103
w
hi
le
it
is
sign
ific
antly
decr
easi
ng
f
or
it
e
m
-
id 1
04 a
nd it
em
-
id 1
02.
In
Ta
ble
2,
we
see
that
it
em
-
id
10
3
retai
ns
the
1
st
ra
nk
with
C
os
ine
sim
il
ari
t
y
as
t
he
m
et
ric
bein
g
use
d,
bu
t
t
her
e
is
a
s
wap
in
the
ra
nk
of
it
em
-
id
104
a
nd
it
em
-
id
102.
H
oweve
r,
t
he
Cosi
ne
si
m
il
arity
com
p
utes
that
the it
em
-
id 1
02
n
ee
ds
t
o be cl
eared
w
it
h ra
nk
2,
w
hile t
he
i
tem
-
id 104 hol
ds
rank
3.
W
it
h
t
he
dataset
,
it
is
evide
nt
that
Pears
on
pe
rfor
m
s
bette
r
w
hen
com
par
e
d
t
o
the
Cosi
ne
.
T
he
Pea
rs
on
si
m
il
arity
has
ta
ken
the
co
nsi
der
at
io
n
of
th
e
pr
e
vious
ye
a
r’
s
sam
e
per
io
d
sal
es
into
ac
count
an
d
giv
e
n
th
e
rankin
g
acco
r
di
ng
ly
,
w
hile
th
e
Cosi
ne
sim
i
l
arit
y
co
ns
ide
rs
the
a
ng
le
s
between
t
he
it
em
s
an
d
j
us
t
f
oc
use
s
on
the
i
m
m
ediat
e
per
i
od
a
nd
dra
ws
the
reco
m
m
end
at
ion.
Alt
hough
not
m
uc
h
di
ff
e
ren
ce
is
there
,
t
he
accu
racy
is
wh
at
m
at
te
rs
f
or the syst
em
.
Fo
r
a
sm
all
dataset
,
the
gr
a
ph
can
be
us
ed
bu
t
the
gr
a
ph
won’
t
be
a
go
od
le
a
rn
e
r
f
or
pr
e
dicti
ng
t
he
fu
t
ur
e
or
tre
nd
s
wh
e
n
c
om
par
ed
t
o
the
j
ob
done
by
a
rec
om
m
end
er
syst
em
.
Also
,
the
grap
h
s
hall
serve
well
with
a
sm
al
le
r
dataset
but
in
tod
ay
'
s
scena
rio
,
the
data
is
dr
a
sti
cal
ly
increas
ing
,
a
nd
the
j
ob
of
t
he
rec
omm
end
e
r
sy
stem
is
to
de
al
with
huge
a
m
ou
nt
of
data
and
le
a
rn
e
ff
ic
i
ently
to
pro
vide
bette
r
rec
omm
end
at
ion
a
nd
accurate
resu
lt
s s
o
t
hat the
people ca
n r
el
y on
su
c
h
sys
tem
s to
overc
om
e losses
by stock accum
ulati
on.
The
ov
e
rall
ou
tc
om
e
po
ints
towa
r
ds
the
it
e
m
-
i
tem
si
m
i
la
rity
reco
m
m
endat
ion
of
t
he
co
ll
abo
rati
ve
filt
ering
sho
wn
belo
w
from
the
dataset
wh
ic
h
has
t
he
m
axim
u
m
nu
m
ber
of
stoc
ks
.
Th
e
it
e
m
s
wh
ic
h
ha
ve
th
e
le
ast
ou
tflo
w
f
r
om
the
inv
ento
ry
has
be
en
ob
t
ai
ned
as
rank
1
fo
ll
owed
by
th
e
it
e
m
s
with
the
m
axi
m
u
m
ou
tflo
w
w
it
h
ra
nk
5.
T
o
s
umm
arize
it,
the
it
em
s
with
a
higher
ra
nk
,
i.e.
r
an
k
1
ne
eds
so
m
e
strat
egies
t
o
be
ad
opte
d
by
the r
et
ai
l s
hops, m
alls,
et
c. s
o t
hat these st
oc
ks
don’t
b
ec
om
e d
ead
-
stoc
k.
Evaluati
on
of
r
esults
ob
ta
ine
d
is
a
n
im
po
rta
nt
crit
erio
n
t
o
know
ho
w
w
el
l
our
syst
em
is
per
f
or
m
ing
,
or
in
oth
e
r
w
ords
to
kn
ow
w
he
ther
the
re
su
lt
s
ob
ta
ine
d
are
correct
or
not.
The
m
os
t
popu
la
r
te
ch
nique
use
d
to
evaluate
the
re
su
lt
s
in
t
he
rec
omm
end
er
syst
e
m
is
preci
sio
n
an
d
recall
.
P
re
ci
sion
(
P)
is
de
f
ined
as
the
num
ber
of
Tr
u
e Po
sit
iv
es
(T
P)
over
th
e
num
ber
of
T
P
pl
us
t
he
nu
m
ber
of
False
P
osi
ti
ves
(FP),
gi
ven
by
t
he
fo
ll
ow
i
ng
(1)
a
nd
(
2).
=
+
(1)
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
Reco
mm
e
nder
syste
m
fo
r s
ur
pl
us
stock
clear
an
ce
(
Vipu
l
Ag
ar
wal
)
3819
Re
cal
l(R)
is
def
ine
d
as
the
num
ber
of
T
rue
Po
sit
ives
(T
P)
ov
e
r
the
nu
m
ber
of
T
P
pl
us
the
nu
m
ber
of
False
Neg
at
ive
s
(
FN),
giv
e
n by the
fo
ll
owin
g (2)
:
=
+
(
2)
Th
us
,
we
us
e
t
he
pr
eci
sio
n
a
nd
recall
s
umm
a
ry
sta
ti
sti
cs
by
cut
-
off
as
s
how
n
in
Table
3
for
Pears
on’s
sim
i
la
rity
m
et
ric.
Si
m
il
ar
ly
,
we
us
e
t
he
pr
eci
sio
n
a
nd
r
ecal
l
su
m
m
ary
sta
ti
sti
cs
by
cu
t
-
off
as
s
how
n
in
Ta
ble
4
f
or
Cosine
si
m
il
arity
m
et
r
ic
.
Table
3.
Pr
eci
s
ion
a
nd
recall
s
umm
ary
us
in
g
P
ears
on
si
m
il
arity
Cu
to
ff
Mean_
p
recisio
n
Mean_
reca
ll
1
1
.0
0
.03
0
4
2
1
.0
0
.06
0
8
3
1
.0
0
.09
1
2
4
1
.0
0
.12
1
6
5
1
.0
0
.15
2
1
6
1
.0
0
.15
2
2
7
1
.0
0
.15
2
4
8
1
.0
0
.15
2
4
9
1
.0
0
.15
2
4
10
1
.0
0
.15
2
5
Table
4.
Pr
eci
s
ion
a
nd
recall
s
umm
ary
us
in
g
C
os
ine
sim
il
arity
Cu
to
ff
Mean_
p
recisio
n
Mean_
reca
ll
1
0
.8
0
.00
4
6
2
0
.8
0
.00
6
8
3
0
.8
0
.00
9
0
4
0
.8
0
.01
6
4
5
0
.8
0
.03
2
2
6
0
.8
0
.03
8
7
7
0
.8
0
.04
1
1
8
0
.8
0
.04
1
1
9
0
.8
0
.04
1
1
10
0
.8
0
.04
1
1
Her
e
the
cut
-
off
ra
nk
ra
nges
from
1
-
10
an
d
f
or
each
ra
nk,
the
m
ean_pr
e
ci
so
n
an
d
m
ea
n_recall
ar
e
giv
e
n.
For
t
he
t
op
-
N
it
em
s
where
we
hav
e
ta
ke
n
for
5
it
em
s,
we
hav
e
the
m
e
an_p
reciso
n
a
nd
m
ean_
recall
giv
e
n
in
the
ta
ble
.
T
o
unde
rstan
d
c
ut
-
off
ra
nk,
we
c
an
say
that
pr
e
ci
sion
or
recall
at
a
giv
e
n
c
ut
-
off
rank,
kee
ping
int
o
account
t
he
res
ults
obta
ine
d
a
re
t
op
-
N
wh
ic
h
a
re
retu
r
ned
by
the
syst
em
,
thu
s
it
is
cal
le
d
as
pr
eci
si
on
at
n
or
P@
n.
We
ha
ve
obtai
ned
the c
ut
-
off
ti
ll
1
0,
w
hich
m
eans th
a
t for the
to
p 10 ra
nk
s
,
the
corr
esp
onding
pr
ec
isi
on
and recall
will
b
e
ob
ta
ine
d.
Fr
om
Table
3,
it
is
see
n
t
hat
f
or
the
Pea
rs
on
si
m
il
arity
m
et
r
ic
,
the
m
ean_
preci
sion
is
1.0
thr
oughout,
wh
il
e
t
he
m
ean_recal
l
inc
rea
ses
sig
nifica
ntly
.
As
e
xp
la
ine
d,
a
m
ean_pr
e
ci
sion
of
1
is
a
ve
ry
good
r
esult.
W
he
n
c
om
par
e
d
to
Table
4
for
cut
off
by
Cos
ine
sim
il
arit
y,
t
he
m
ean_
preci
sion
is
0.8
w
hich
is
lo
wer
tha
n
the
m
ean_
preci
sio
n
obta
ined
th
rough
the
Pear
son
sim
il
arity
m
e
tric
.
Sim
il
arl
y,
we
see
the
re
ca
ll
has
dec
rease
d
f
or
the
Cosi
ne
m
etr
ic
whe
n
c
om
par
ed
w
it
h t
he P
earson m
et
ric.
Higher
t
he
valu
e
of
preci
sio
n,
i
t
giv
es
m
os
t
of
the
predict
ed
la
bels
as
c
orrect,
wh
il
e
lo
w
r
ecal
l
ind
ic
at
es
that m
os
t of it
s pred
ic
te
d
la
be
ls are i
nc
orr
ect
. S
yst
em
s w
it
h hig
h recal
l b
ut
low p
recisi
on
r
et
urns
m
any re
su
lt
s,
bu
t
m
os
t
of
it
s
pr
e
dicte
d
la
bel
s
are
i
ncorr
ect
wh
e
n
c
om
par
e
d
t
o
the
trai
ni
ng
la
bels.
A
syst
e
m
with
hi
gh
preci
sion
bu
t
t
he
l
ow
rec
al
l
is
just
the
opposit
e
wh
ic
h
will
retu
rn
res
ults,
wh
e
re
m
os
t
of
it
s
pr
e
dicte
d
la
bels
are
c
orrec
t
wh
e
n
c
om
par
e
d of t
he
trai
ning labels
.
In
an
in
form
ati
on
retrie
val
sy
stem
,
a
preci
sion
sc
or
e
of
perfect
1.0
in
dicat
es
that
e
ver
y
r
esult
w
hic
h
was
obta
ine
d
by
the
searc
h
w
as
rele
va
nt
bu
t
does
n’
t
say
a
ny
thing
ab
out
w
h
et
he
r
al
l
t
he
r
el
evan
t
inf
orm
at
ion
was
ret
rieve
d
or
not.
On
the
c
on
t
rar
y,
a
recal
l
scor
e
of
pe
rf
e
ct
1.0
in
dicat
es
that
al
l
the
res
ults
obta
ined
by
the
search
w
a
s r
el
e
van
t
but d
oes
not say
how
m
any irrele
van
t
r
e
su
lt
s w
e
re als
o ob
ta
ine
d.
To
c
on
cl
ud
e
,
P
earson
’s
sim
i
lar
it
y
gav
e
bette
r
res
ults
w
hen
com
par
ed
with
the
Cosine
sim
il
arity
by
hav
in
g
a
bette
r pr
eci
sio
n
a
nd r
ecal
l
value ov
e
r
t
he on
e
obtai
ned b
y t
he
Cosine
sim
i
l
arit
y.
Hen
ce
,
the
syst
e
m
giv
es
a
perfect
sc
or
e
of
preci
sio
n
i.e.
1.0
an
d
ind
ic
at
e
al
l
the
rele
va
nt
reco
m
m
end
at
ion
s
a
re
obtai
ne
d. T
he recal
l
va
lue ca
n be
inc
reased
bu
t
it
w
ou
l
d
le
a
d t
o a
s
ign
ific
a
nt
decre
ase in
the preci
sio
n v
al
ue.
I
t i
s
a cha
ll
eng
e
for
the
re
searche
rs
t
o h
ave a hig
h valu
e o
f
both
the
preci
sion
a
nd r
ec
al
l at
the sam
e tim
e, b
ut it
’s
sti
ll
n
ot
achieve
d an
d
t
hu
s
it
leaves
be
hind a
sco
pe fo
r
im
pr
ov
em
ent in e
ver
y
dom
a
in.
7.
CONCL
US
I
O
N
Re
com
m
end
at
ion
syst
em
has
pro
ved
to
be
of
great
sup
port
i
n
va
rio
us
dom
a
ins
t
o
filt
er
out
the
c
onte
nt
and
pro
vid
e
use
fu
l
rec
omm
e
nd
at
io
ns
m
aking
the
w
ork
of
t
h
e
us
e
rs
eas
ie
r.
In
t
his
pa
per,
t
he
desi
gn
of
a
reco
m
m
end
er
s
yst
e
m
fo
r
the
s
urplus
stoc
k
is
descr
i
bed
w
hic
h
will
help
t
he
retai
le
rs
in
i
de
ntifyi
ng
the
it
e
m
that
can
le
a
d
t
o
the
accum
ulati
on
of
st
ock
s
.
T
his
m
et
ho
d
im
ple
m
ents
an
it
e
m
-
it
e
m
colla
bo
rat
ive
filt
er
in
g
te
chn
i
que
,
with
the
Pea
rs
on
sim
il
arity
m
et
ric
and
t
he
Cosine
sim
il
arity
m
et
ric
wh
ic
h
will
be
us
e
d
f
or
ide
ntifyi
ng
the
to
p
-
N
it
em
s
as
m
entione
d
in
th
e
r
esults
of
t
he
sim
il
ar
cat
egory
wh
ic
h
ha
ve
the
hi
ghest
st
ock
accum
ulati
on
s
ta
ti
ng
rank
1
in
the
i
nv
e
nt
ory
over
a
pe
rio
d
an
d
gi
ve
a
n
e
ff
ic
ie
nt
rec
omm
end
at
ion
ab
out
t
he
i
tem
-
id
to
cl
ear
the
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.
9
, N
o.
5
,
Oct
ober
201
9
:
3813
-
3821
3820
existi
ng
stoc
k. The
pr
e
ci
sion and
recall
o
bta
ined
by
t
he
Pe
arson
a
nd
t
he
Cosine
sim
il
ari
ty
m
et
ric
al
so
ind
ic
at
e
the
perform
ance
of
Pears
on
sim
il
arity
m
et
ric
trum
ps
over
t
hat
of
the
Cosi
ne
sim
il
arity
m
et
ric.
The
pr
eci
sio
n
and
recall
also
ind
ic
at
e t
he pe
rfor
m
ance
of t
he sy
stem
and
there
is al
ways a
sco
pe
to
enh
ance t
he
p
e
rform
ance
wh
ic
h
is
a
ch
al
le
ng
e
f
or
t
he
researc
hers.
Ex
per
ie
nc
es
of
this
resea
rc
h,
pro
blem
s,
and
com
e
back
for
the
pro
blem
s
has
been
s
ha
red
al
ong
wit
h
t
he
r
esults
s
o
that
the
resea
rch
e
rs
don’t
face
a
ny
tro
ub
le
in
dupl
ic
at
ing
the
sam
e
or
cr
eat
ing
a
ne
w
s
yst
e
m
of
a
si
m
il
ar
kind.
T
his,
in
tur
n,
broa
de
ns
the
sc
ope
of
researc
he
rs
to
le
arn
m
or
e
and
over
com
e
the
exis
ti
ng
pro
blem
s,
identify
in
g
do
m
ai
ns
wh
e
re
r
ecom
m
end
er
s
yst
e
m
s
hav
e
not
bee
n
i
m
ple
m
ented.
Along
with
t
ha
t,
a
c
om
par
ison
ca
n
be
m
ade
with
ot
her
te
chn
iq
ues
a
nd
the
best
one
can
be
discusse
d
a
nd
i
m
ple
m
ented
as
a
fu
tu
re
of
t
he
existi
ng
w
ork
done.
T
hus,
on
e
ca
n
creat
e
a
reco
m
m
en
datio
n
syst
e
m
wh
ic
h
will
be
m
or
e
ef
fici
ent
in
ha
ndli
ng
the
iss
ues
a
nd
kill
the
m
anu
al
work
pr
ess
ur
e
t
ha
n
t
ho
s
e
existi
ng
in the p
resen
t
s
yst
e
m
.
ACKN
OWLE
DGE
MENTS
We
would
li
ke
to
tha
nk
the
Al
m
igh
ty
for t
hr
owin
g l
ig
ht
th
rough
ou
t
our
to
ugh
phases
w
hic
h
we
cam
e
acro
s
s
du
rin
g
our
wor
k,
with
ou
t
his
blessin
gs
,
it
wouldn
’t
ha
ve
been
pos
sible
to
ac
hiev
e
it
.
We
e
xten
d
war
m
gr
at
it
ude
t
o
our
sup
porters
a
nd
m
otivators
f
or
hav
i
ng
c
onfide
nce
i
n
us
a
nd
pro
vid
i
ng
un
e
nd
ing
sup
port
i
n
e
ver
y
po
s
sible
way
t
o
overc
om
e
al
l
the
hurd
le
s
in
this
wor
k
a
nd
st
and
ou
t.
Last
ly
,
we
would
li
ke
to
t
ha
nk
ou
r
pa
ren
ts
for
ha
ving
tr
us
t
in
us
a
nd
pro
vid
in
g
al
l
the
enc
oura
gem
ent
and
m
otivati
on
re
qu
i
red
t
o
acco
m
pl
ish
the
obje
ct
ive
.
REFERE
NCE
S
[1]
F. Is
inkay
e
,
Y. F
ola
ji
m
i
,
and B
.
Ojokoh,
“
Rec
om
m
enda
ti
on
s
y
ste
m
s:
Princi
ple
s,
m
et
hods
and
evalua
ti
on
,
”
Eg
ypt
i
an
Informatic
s J
our
nal
,
vo
l. 16, no.
3,
pp
.
261
-
273
,
2015.
[2]
C.
He,
D.
Parr
a,
and
K.
Ve
rbe
rt
,
“
Inte
ra
ctive
r
e
comm
ende
r
s
y
st
ems
:
A
surve
y
of
the
st
at
e
of
t
he
art
and
fu
tur
e
rese
arc
h
ch
al
l
en
ges
and
opp
ortu
nit
ie
s
,
”
Ex
p
ert
S
yste
ms
wit
h
Applicat
ions
,
vol
.
56
,
pp.
9
-
27,
2016.
[3]
J.
B.
Sch
afe
r
,
J.
Kons
ta
n,
an
d
J
.
Ri
edi
,
“
Rec
om
m
ende
r
s
y
st
ems
in
e
-
comm
erc
e
,
”
Proc
ee
d
ings
o
f
the
1st
ACM
conf
ere
n
ce on
E
l
ec
troni
c
commer
ce
-
EC
99
,
pp.
1
58
-
166,
1999
.
[4]
S.
Alam,
“
Intelligent
web
usage
cl
uster
ing
b
ase
d
re
comm
ende
r
s
y
stem,
”
Proce
ed
ings
of
th
e
fi
f
th
ACM
conf
ere
n
ce
on
Recom
mende
r syste
ms
-
R
ec
S
ys
11
,
pp.
367
-
3
70,
2011
.
[5]
S.
Alam,
G.
Do
bbie
,
and
P.
Rid
dle
,
“
Parti
c
le
Sw
arm
Optimiza
t
ion
Based
Clust
eri
ng
of
W
eb
U
sage
Dat
a,”
200
8
IEE
E
/WIC/
ACM
Int
ernati
onal
C
onfe
renc
e
on
W
e
b
Int
el
l
ige
nc
e
an
d
Int
el
l
ige
nt
Agent
Te
chnol
ogy
,
p
p.
451
-
454
,
2008
.
[6]
J.
Cl
eland
-
Huan
g
and
B
.
Mobasher,
“
Us
ing
da
ta
m
ini
ng
and
r
ecom
m
ende
r
s
y
st
e
m
s
to
sc
al
e
up
t
he
r
equi
r
ements
proc
ess,”
Procee
dings
of
th
e
2nd
int
ernati
ona
l
wo
rkshop
on
Ultra
-
large
-
scale
soft
ware
-
int
ensi
ve
systems
-
UL
SSIS
08
,
pp
.
3
-
6
,
200
8.
[7]
Y.
H.
Cho,
J.
K.
Kim
,
and
S.
H.
K
im,
“
A
per
sonalize
d
re
comm
ende
r
s
y
stem
base
d
o
n
web
usage
m
in
ing
and
de
ci
sion
tre
e
induction
,
”
Ex
pert
Syste
ms
wit
h
App
licati
on
s
,
vol. 23, no. 3,
pp.
329
-
342
,
20
02.
[8]
S.
Sivapa
l
an,
A.
Sadeghi
an
,
H.
R
ahna
m
a,
and
A.
M.
Madni,
“
Recom
m
ende
r
sy
st
e
m
s
in
e
-
comm
er
ce
,
”
2014
Worl
d
Aut
omation
Con
gress
(
WAC
)
,
pp.
179
-
184
,
2014
.
[9]
P.
Barga
h
and
N
.
Mishra,
“
Solving
Sparsit
y
Problem i
n
Movie
Ba
sed
rec
om
m
endation
s
y
st
em,”
Ad
vanc
es
in
Image
and
Vi
d
eo Proces
sing
,
vol. 4, no. 3, pp. 25
-
31,
20
16.
[10]
K.R.
B
indu,
Rh
ama
L
al
gudi
Vis
wesw
ara
n,
P.C.
Sachi
n,
Kundav
ai
Devi
Solai
an
d
Soundar
y
a
Gu
nase
kar
an
,
K.
R.
Bindu,
R.
L
.
Visw
eswara
n,
P.
C
.
Sachi
n,
K.
D.
So
la
i
,
and
S.
Gun
ase
kar
an
,
“
Reduc
i
ng
th
e
Co
ld
-
Us
er
and
Cold
-
I
te
m
Problem
in
R
ecom
m
ende
r
S
y
ste
m
b
y
Redu
ci
ng
the
Spa
rsi
t
y
of
t
he
Sparse
Matr
i
x
and
Address
in
g
the
Div
ersity
-
Acc
ura
c
y
Probl
e
m
,
”
Ad
vanc
es
in
Inte
lligen
t
Syst
ems
and
Compu
ti
ng
Proc
ee
ding
s
of
Inte
rnation
al
Confe
ren
ce
o
n
Comm
unic
ati
on
and
Net
works
,
p
p.
561
-
570
,
201
7.
BIOGR
AP
HI
ES OF
A
UTH
ORS
Vipu
l
Agar
w
al
is
a
postgr
adu
at
e
studen
t
in
Master
s
of
Com
pute
r
Appli
ca
t
ion
a
t
CHRIS
T
(Dee
m
ed
to
be
Univer
sit
y
),
Ban
gal
ore
.
M
y
ar
eas
of
int
er
ests
ar
e
Data
Anal
y
t
ics
and
Mac
hin
e
Le
arn
ing.
Ap
art
from
the
se,
m
y
m
aj
or w
ork and
proje
c
ts
are
in r
e
comm
ende
r s
y
st
ems
and pattern
rec
ogni
ti
on.
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
Reco
mm
e
nder
syste
m
fo
r s
ur
pl
us
stock
clear
an
ce
(
Vipu
l
Ag
ar
wal
)
3821
Vijay
alaks
hmi
A,
works
as
an
As
sistant
profe
ss
or
at
CHRIS
T
(De
emed
to
be
Univ
ersity
)
,
Banga
lor
e.
She
has
completed
h
er
Ph.D.
in
fa
ce re
cognition.
Her
fie
lds
of
in
te
rest
inc
lud
e
pa
ttern
rec
ogni
ti
on,
Ma
c
hine
learni
ng
,
an
d
Artif
icial
In
te
l
l
ige
nc
e.
Curr
ent
l
y
,
sh
e
is
do
ing
a
proje
c
t
on
IoT
and
m
ac
hin
e lea
rning.
Evaluation Warning : The document was created with Spire.PDF for Python.