Int
ern
at
i
onal
Journ
al of Inf
orm
at
ic
s
and
Co
m
munic
at
i
on
Tec
hn
olog
y (IJ
-
I
CT)
Vo
l.
8
,
No.
3
,
D
ecembe
r
201
9
, pp.
128
~
138
IS
S
N:
22
52
-
8776
, DO
I: 10
.11
591/ij
ic
t.v8
i
3.p
p128
-
138
128
Journ
al h
om
e
page
:
http:
//
ia
escore.c
om/j
ourn
als/i
ndex.
ph
p/IJI
C
T
Inventory
predicti
on and
mana
gement in
Nigeri
a using m
ar
k
et
basket a
na
l
ysis asso
ciativ
e rule mi
ning: me
metic al
gorithm
based a
ppro
ach
Arn
old
Ad
im
ab
u
a Oju
go
1
,
An
dre
w Ok
onji E
boka
2
1
Depa
rt
m
ent
of
Mathe
m
at
i
cs/Co
mput
er
Scie
n
ce
,
Feder
al Unive
rsi
ty
of
Pe
troleum Resource
s E
f
fur
un,
Niger
ia
2
Depa
rt
m
ent
of
Comput
er
Scie
n
ce
Educat
ion
,
Fe
der
al Col
l
ege of Educ
a
ti
on
(
Tech
nic
a
l)
As
aba,
Ni
ger
ia
Art
ic
le
In
f
o
ABSTR
A
CT
Art
ic
le
history:
Re
cei
ved
J
ul
12, 2
01
9
Re
vised
N
ov
13
, 2
01
9
Accepte
d Nov
2
8
, 201
9
A
well
-
p
rep
ar
ed
abstr
act
ena
b
le
s
th
e
re
ade
r
to
id
ent
ify
th
e
basi
c
cont
en
t
o
f
a
docum
en
t
quic
k
l
y
and
accuratel
y
,
to
d
et
e
rmi
n
e
it
s
rel
ev
ance
to
th
eir
in
te
r
ests,
and
thus
to
d
eci
de
whe
the
r
to
r
e
ad
the
docu
me
nt
in
i
ts
ent
ir
et
y
.
T
he
Abs
t
ra
ct
should
be
infor
ma
ti
v
e
and
co
mpl
etely
se
lf
-
ex
pla
na
tory,
prov
i
de
a
cl
e
ar
stat
ement
of
the
problem,
th
e
pr
oposed
appr
o
ach
or
solu
ti
on
,
a
nd
point
out
ma
jor
f
indi
ngs
a
nd
conc
lusions.
The
Abs
tra
c
t
sh
ould
be
100
to
2
00
words
in
le
ngth
.
Th
e
abstr
ac
t
should
be
wri
tt
en
in
th
e
p
ast
t
e
nse.
Stand
ard
no
me
nc
la
tur
e
should
be
used
a
nd
abbr
ev
ia
t
ions
should
be
avoi
d
ed.
No
li
t
erature
should
be
ci
t
ed.
Th
e
keyw
ord
li
st
provid
es
the
opportun
it
y
t
o
add
keywords,
used
by
the
inde
xing
and
ab
strac
t
ing
service
s,
in
addition
to
th
ose
al
r
ea
dy
pr
ese
nt
in
the
ti
tle.
Judic
ious
u
se
of
k
eywords
ma
y
inc
r
ea
se
th
e
e
ase
with
whi
c
h
intere
st
ed
par
ties c
an
lo
cat
e
our art
ic
l
e.
Ke
yw
or
d
s
:
Associ
at
ive rul
e minin
g
Con
ce
pt
dr
ift
Data mini
ng
M
ar
ket
bas
ket
Pr
e
dicti
on
a
nal
ys
is
Sh
el
ve
p
la
ce
m
ent
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
:
Arno
l
d Ad
i
ma
bu
a
O
j
ugo
,
Dep
a
rtme
nt of
M
at
he
mati
cs/
Com
pu
te
r
Scie
nc
e
,
Fede
ral U
niv
e
r
sit
y
of Pet
ro
le
um R
eso
urces E
ffuru
n,
P.M.B
1221, E
ffuru
n,
Warri,
Delt
a Stat
e,
Niger
ia
.
Emai
l:
mar
yarno
l
do
j
ugo@g
mail
.co
m
1.
INTROD
U
CTION
The
w
or
ld
ove
r,
businesse
s
ta
ke
st
ock
an
d
i
nv
e
ntorie
s
of
their
daily
pro
duct
ion
s
o
as
t
o
acco
un
t
f
or
goods
a
nd
se
rvi
ces
re
nd
e
red
t
o
thei
r
cl
ie
nt
i
n
e
xchan
ge
for
mone
y.
I
nv
e
nt
or
ie
s
ha
ve
bee
n
viewe
d
by
m
any
as
raw
materia
ls, w
ork
in p
r
ogre
ss
or
fi
nish
e
d
pro
duct
s
that
ar
e
store
d
to
mee
t
the
sup
ply
d
e
man
ds
o
f
c
ons
um
e
rs
[1,
2].
I
f
the
a
moun
t
of
in
ve
ntory
is
le
ss
th
an
am
ount
of
a
ct
ual
nee
d,
t
he
bu
si
ness
ma
y
lose
th
e
op
port
un
it
y
to
maximize
sal
es.
T
hey
ma
y
al
so
l
os
e
pote
ntial
cl
ie
nts,
l
os
e
l
oyal
ti
es
as
well
as
lose
antic
ipa
te
d
ma
xim
um
prof
it
s;
Wh
il
e,
if
t
hey
stock
to
o
m
uc
h
of
the
i
nvent
ory,
it
will
incr
ease
the
c
os
t
of
mai
ntena
nce
and
sto
rage,
a
nd
al
s
o
conseq
ue
ntly
–
re
du
ce
the
pro
fi
t mar
gin
s
[3,
4].
Inve
nto
r
y
s
uppl
y
de
man
d
valu
e
chain
a
nd
it
s
mana
geme
nt
ha
ve
ri
pp
le
d
ma
ny
busines
ses
with
a
ra
ng
e
of
c
ompli
cat
ion
s.
T
hus,
the
fi
el
d
has
at
tract
e
d
the
at
te
ntio
n
of
m
an
y
rese
ar
cher
s
a
nd
prac
ti
ti
on
ers
no
wa
days.
[5]
us
e
d
movin
g
a
ver
a
ge
m
o
de
l
for
a
c
ompa
ny
wit
h
fluctua
ti
ng
dema
nd,
pro
ve
d
that
m
oving
ave
rage
is
able
t
o
accomm
odat
e
rap
i
d
cha
nges
in
data;
A
nd
quit
e
su
it
able
f
or
c
ompanies
with
co
ndit
ions
of
high
va
ri
et
y
of
pro
du
ct
s
a
nd
r
aw
mate
rial
s.
This
met
hod
is
ap
propriat
e
w
hen
us
e
d
t
o
pre
d
ic
t
long
-
te
rm
pr
e
dicti
on
s
.
T
hus,
[5]
exp
l
or
e
d
ot
her
stu
dies
t
hat
e
mp
lo
ye
d
e
xponentia
l
s
mooth
ing
m
od
el
[
6]
and
the
B
ox
-
Je
nk
i
ns
aut
o
-
regressi
ve
integrate
d
mov
ing
ave
rag
e
[
7]
.
T
he
i
nh
e
ren
t
li
mit
at
ion
s
of
e
ach
meth
od
an
d
oth
e
rs,
acco
unte
d
f
or
the
dif
ficult
y
in
ap
ply
i
ng
th
e
meth
od
to
knowle
dge
-
base
.
H
oweve
r,
[
8]
empl
oy
e
d
ge
ne
ti
c
al
go
rith
m
to
pre
dict
inv
e
ntor
y
stocks
a
nd
pro
ved
th
at
me
meti
c
al
gorithm
offer
s
ma
ny
be
ne
fits
s
uch
as
it
s
bein
g
a
c
omp
utati
on
al
more
ef
fici
en
t
al
gorithm,
m
ore accu
rate a
nd
le
ss ti
me
-
c
on
s
um
in
g. I
n f
ur
t
he
ran
ce
, [9
]
e
xt
end
e
d t
his
w
ork usin
g dee
p l
earn
i
ng
and
note
d
that
there
are
i
nh
e
r
ent
chall
en
ges
in
us
i
ng
ge
netic
al
gorithm.
T
hat
w
hile,
arti
f
ic
ia
l
neu
ral
net
works
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Inf & C
ommu
n Tec
hn
ol
IS
S
N:
22
52
-
8776
In
ve
nto
ry
pred
ic
ti
on
an
d man
ag
e
me
nt in
Ni
ge
ria
us
in
g mar
ket
b
ask
et
an
alysis …
(
Ar
nold
Ad
i
mab
ua O
j
ugo
)
129
(ANN) ar
e
su
it
ed
f
or
lea
rn
i
ng the und
e
rlyi
ng p
r
obabili
ti
es o
f
feat
s
of
inte
re
st i
n
mar
ket b
a
sk
et
an
al
ys
is
–
deep
le
arn
in
g
is
bes
t
su
it
ed
to
pr
e
di
ct
the
amo
un
t
of
pro
du
ct
i
nvento
ry
nee
ds
due
to
oth
e
r
s
pa
ti
al
data
therein
li
ke
pr
e
dicti
on
of
c
on
ce
pt
e
vo
l
ution,
c
on
ce
pt
dr
i
ft
am
ongs
t
othe
rs.
Th
us
,
the
r
at
ion
al
e
a
nd
m
ai
ns
ta
y
f
oc
us
f
or
us
i
ng
memet
ic
alg
or
i
thm
her
e
is to
adv
a
nce t
he us
e of
ANN
i
n pre
dicti
on
of in
ve
ntory
as
her
e
pro
po
se
d.
M
BA
is
a
data
minin
g
met
hod,
f
ocusi
ng
on
i
den
ti
ficat
io
n
of
pro
du
ct
s
t
hat
a
re
pu
rch
ase
d
at
same
ti
me
on
each
tra
ns
a
ct
ion
[10
]
.
O
utp
ut
of
M
BA
is
a
set
of
r
ules
t
hat
in
dicat
e
th
e
pro
duct
s
that
are
purc
hase
d
on
the
same
ti
me.
T
his
outp
ut
will
be
us
e
d
as
in
pu
t
f
or
the
pr
e
dicti
on
of
i
nv
e
nt
ory.
The
ru
le
s
ge
ner
at
e
d
by
M
B
A
a
r
e
associat
ion
r
ule(s)
a
re
of
the
f
orm:
If
a
nteced
ent
(A),
t
hen
c
on
s
eq
ue
nt
(B).
Eac
h
r
ule
is
eq
uipped
with
a
s
uppor
t
le
vel
that
i
nd
ic
at
es
the
num
be
r
of
tra
ns
act
io
ns
c
onta
inin
g
A
a
nd
B
a
nd
c
onfide
nce
le
ve
l
that
is
a
mea
su
re
of
accurac
y
wh
ic
h
is
the
r
ule
of
associat
ion
r
ul
es.
[
11,
12]
eac
h
r
ule
is
al
so
e
qu
i
pp
e
d
with
a
n
e
xp
ect
e
d
c
onditi
on
a
nd
a
li
ft
so
t
ha
t
for
each
ant
ecedent
(A)
a
nd
co
ns
e
quence
(B),
t
he
s
uppo
rt,
co
nfi
de
nce,
exp
ect
e
d
c
onfi
den
ce
,
and
li
ft
a
re
as
in
E
qu
at
i
on
s
be
low
an
d
fig
.
1
(Yus
uf,
Pr
at
ikto
an
d
Ge
rry,
20
06).
S
uppo
se
we
def
ine
h
⊂
{1
,
2
,…
P} a
nd A,
B
⊂
h
–
w
her
e
A U
B =
h
, a
nd
A ∩ B =
⌽
. T
hen,
we have
that:
(
ℎ
)
=
(
,
)
=
.
&
.
(1)
Eq. 1
. P(h) is
pre
valence
or
s
uppo
rt
w
hich
yields
how
often the
combi
na
ti
on
A
a
nd B c
o
-
occ
ur
s
.
(
|
)
=
(
ℎ
)
(
)
=
.
&
.
(2)
Eq.
2.
P(
B|
A
)
is
confi
dence
va
lue
w
hich
yields
the
c
onfi
dence
that
it
em
B
app
ea
rs
in
the
bas
ket
giv
e
n
A
is al
rea
dy in
the b
a
sk
et
.
T
hus,
we
use t
he
rule A
→
B.
(
|
)
=
(
)
(
ℎ
)
=
.
(3)
Eq.
3.
P
(A)P
(B
)
ex
pected
c
onfidence
yields
the
co
nf
i
den
ce
on
how
fr
e
qu
e
nt
it
ems
A
an
d
it
ems
B
co
-
occur i
n
the
nu
mb
e
r of
ti
mes
that the it
ems
B
is ch
o
se
n
a
nd
placed i
n
t
he b
asket.
=
(4)
Finall
y,
(
4
)
.
L
(
A,
B
)
li
ft
of
t
he
r
ule
A
→
B
yields
a
measu
re
of
how
m
uc
h
more
co
nf
i
de
nt
we
a
re
i
n
it
em B g
ive
n
t
hat w
e
see it
e
m A in t
he bas
ket.
Figure
1. Sc
he
mati
c represe
ntati
on
of a rule
M
BA
is
a
s
ubs
et
of
ma
rk
et
re
search
t
hat
ma
ny
resear
che
rs
are
cu
rr
e
ntly
pa
ying
sp
e
ci
al
at
te
ntion
to
with
m
or
e
deta
il
ed
in
[
3].
Ta
n
g,
K
,
et
al
[
13]
pro
po
se
d
a
n
a
ppr
oac
h
to
perform
ma
rket
bas
ke
t
analysis
in
a
mu
l
ti
-
store,
m
ulti
-
pe
rio
d e
nviro
nm
e
nt.
Ch
en,
Y
,
et
al
[
14]
note
d
th
at
most
m
odel
s
us
e
d
in
deali
ng
with
ma
rk
et
bas
ket
pro
blem
c
ou
l
d
not
disc
over
any
imp
ort
ant
purc
hasin
g
pat
te
rn
s
w
hen
a
nd
wh
e
re
mu
lt
iple
sto
res
e
xis
t.
The
y
dev
el
op
e
d
a
m
et
hod
t
o
overc
om
e
this
we
a
kness;
w
hile,
Y
un,
C
,
et
al
cl
ust
ered
data
of
mar
ket
bask
et
us
in
g
a
novel
meas
ur
e
ment
t
hey
nam
ed
cat
e
gor
y
-
ba
sed
a
dhere
nce
[15]
.
C
avi
que
,
L
c
onve
rted
m
ark
et
bas
ket
pr
ob
le
m
into
a
ma
ximum
-
weig
hted
cl
iqu
e
pro
blem
f
or
disco
ver
i
ng
la
rg
e
it
em
set
patte
rn
s
[6
]
.
A
ccordin
g
to
[
16]
the
y
dev
el
op
e
d
opti
miza
ti
on
m
ode
l
fo
r
sh
el
f
-
s
pac
e
mana
geme
nt
pro
blem
in
w
hich
pr
oducts
a
re
gro
up
e
d
as
fa
mil
ie
s
and
the
loc
at
io
n
of
eac
h
famil
y
is
determin
ed
on
the
s
helf
li
ke
cat
al
ogin
g.
T
hey
c
onsidere
d
she
lf
l
ocati
on
eff
ect
on
sal
es;
but
,
di
d
not
at
te
nd
the
cr
os
s
-
sel
li
ng
ef
fect.
Th
us
,
the
y
di
d
not
use
the
purc
has
e
data.
Ni
ero
p,
E
,
et
a
l
[17]
pro
pose
d
a
meth
od
f
or
de
al
ing
with
s
he
lf
-
s
pace
ma
na
geme
nt
pro
ble
m
that
co
ns
ist
s
of
t
wo
-
par
ts.
In
t
he
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
In
t J
Inf & C
ommu
n Tec
hn
ol
,
V
ol.
8
,
No.
3,
Dec
201
9
:
12
8
–
138
130
first
ph
a
se,
a
st
at
ist
ic
al
mo
del
was
prov
i
de
d
to
mea
sure
t
he
impact
of
s
helf
la
yout
on
sal
es.
In
the
sec
ond
part
simulat
ed
a
nne
al
ing
(
SA)
was
us
e
d
to
maxi
m
iz
e
exp
ect
e
d
to
ta
l
pr
ofi
t.
T
hey
al
so
li
ke
[16],
did
not
co
ns
ide
r
the
associat
ion r
ul
es fro
m c
us
to
mers
’ purcha
sing data.
T
hu
s
, did
not u
se it
to
ma
xim
iz
e c
r
os
s
-
sel
li
ng
e
ff
e
ct
.
Re
cent
re
searc
hes
c
onside
r
ot
her
pr
ob
le
m
s.
Accor
ding
t
o
S
ara
f,
R.
and
Patil
,
S
the
y
pr
opose
d
a
bo
tt
om
-
up
hie
rar
c
hical
cl
us
te
r
-
m
od
el
for c
lu
ste
rin
g
r
et
ai
l
it
ems
[18]
. T
o
do
this,
th
ey a
pp
li
e
d t
he
con
ce
pt
of
‘
dis
ta
nce’
betwee
n
t
he
e
ntit
ie
s
or,
gro
ups
of
e
ntit
ie
s
to
ac
hieve
the
pur
po
se
of
ma
rk
et
-
bas
ket
a
na
lysis.
M
a
rk
et
bas
ket
analysis
is
no
w
em
ployed
by
ma
ny
resear
cher
s
t
o
ot
her
app
li
cable
ta
s
ks.
Shiokawa,
Y
,
et
al
app
li
ed
mar
ket
bas
ket
an
al
ys
is
f
rame
work
to
visu
al
iz
e
tra
nsa
ct
ion
dat
a
t
o
assess
t
he
var
i
ou
s
human
li
fe
styles
[
19]
.
Sol
net
,
D
,
et
al
al
so
st
ud
i
ed
pote
ntial
s
t
o
gro
w
hote
l
r
evenue
by
e
xp
l
or
i
ng
m
os
t
at
tract
ive
se
rv
ic
es
an
d
pr
oducts
that
ca
n
at
tract
/sa
ti
sfy
gu
e
sts
a
nd
e
nc
oura
ge
t
hem
to
re
peat
t
heir
pu
rch
ase
[20]
.
I
n
furthe
ran
c
e
,
[5]
e
xp
l
or
e
the
c
ultur
al
beh
a
viou
r
of
c
on
s
ume
rs
. F
ur
t
her
stu
dies
ca
n
ref
e
r t
o [
21
-
26
].
T
he
re
viewin
g
of
relat
ed
res
earches
re
veal
that
a
main
f
ocus
of
ma
rk
et
bask
e
t
anal
ys
is
a
nd
it
s
a
pp
li
cat
io
n
–
is
gea
red
to
wards
creat
ing
a
m
or
e
ef
fici
en
t
op
ti
miza
ti
on
al
gorithm
f
or
dat
a
minin
g.
We
c
an
a
pp
l
y
e
vo
l
ution
a
ry
model
a
nd
ass
ociat
ion
ru
le
mi
ning
[27
-
32].
2.
MEMETI
C
B
AY
ES
IA
N NE
TWOR
K EX
PERI
MENTA
L FR
AM
E
W
ORK
Ev
olu
ti
onar
y
a
lgorit
hm
seeks
to
e
xploit
his
toric
numeric
data
a
nd
ex
plore
human
kn
owle
dge
vi
a
mathemat
ic
m
od
el
s
an
d
sym
bo
li
c
reas
on
i
ng
to
yield
a
n
ou
t
pu
t
that
is
toler
ant
to
im
pr
eci
s
ion
,
no
ise
,
un
c
ertai
nt
y
and
pa
rtia
l
truth
as
a
pp
li
ed
t
o
it
s
inp
ut
[
33]
.
I
t
evo
lve
s
into meta
-
r
ules
f
or
co
nst
raint
sat
isfact
ion
ta
s
ks
th
at
use
intel
li
gen
t
age
nts
in
vect
or
s
pace
to
see
k
f
or
opti
mali
ty.
These
al
go
rith
ms/mo
dels
are
insp
ire
d
by
e
voluti
on,
beh
a
vioral
patte
rn
s
in
bi
ologica
l
popula
ti
on
s
an
d
natu
r
e
la
w
s
to
mi
mic
a
ge
nts
see
king
s
ur
viv
al
[34,
35]
a
s
the
y
hav
e
pro
ven
e
f
fici
ent
in
c
omplex
op
ti
miza
ti
on.
Simpl
y
put
,
e
vo
l
utio
nary
model
at
te
m
pts
to
e
xp
l
or
e
dynamic
processes
t
hro
ugh
e
xploit
at
ion
of
obse
rv
e
d
data
t
o
yield
an
outp
ut
t
hat
ex
hib
it
s
r
obus
t
ness,
c
on
ti
nu
ous
adap
ta
ti
on
an
d
flexibili
ty
–
w
hi
le
disp
la
ying
t
he
un
der
l
ying
pro
bab
il
it
ie
s
of
data
feats
of
in
te
rest.
Th
us
,
it
seeks
an
outp
ut
feat
with
unc
on
t
ro
l
la
ble
const
rain
ts
modele
d
within
the
m
odel
s
input
that
m
ay
not
be
e
xpli
ci
tl
y
pr
ese
nt
in
the
s
earch
sp
ace
bu
t confine
d
t
o re
al
par
a
mete
rs
a
s w
el
l as li
mit
e
d by bo
unda
ry
values
[36
-
38].
2.1
.
Ar
tifici
al
neural
n
et
w
or
k (ANN
)
A
N
N
data
pr
ocessin
g
m
od
el
is
insp
ire
d
by
ne
uro
ns
in
the
huma
n
br
ai
n.
T
hus,
consi
sts
of
interco
nnect
ed
ne
uro
ns
(
node
s)
with
capa
bili
ty
to
le
ar
n
by
exam
ple
that
makes
the
m
un
iversal
e
sti
mators
.
As
it
processes
da
ta
,
it
s
node
s
s
har
es
data
sig
na
ls
an
d
a
dju
st
it
s
weig
hts
a
nd
bias
a
dju
stm
ents
re
presenti
ng
the
sy
na
ps
e
ax
ons
an
d
dendr
it
es
to
in
dicat
e
co
nnect
ion
stre
ng
t
h
betwee
n
s
ynapses
res
pecti
ve
ly
[
38
-
39].
Sign
al
s
are
co
nverte
d
s
o
that
weig
hts
are
adj
us
te
d
as
le
arn
in
g
occ
urs
an
d
is
s
umm
ed
by
a
n
adde
r
.
De
pe
nd
i
ng
on
ta
sk,
it
s
act
ivati
on
fun
ct
io
n
li
mit
s
i
ts
outp
ut
[
40
-
41]
to mod
ulate
associat
ed
i
np
uts
a
nd
no
nline
ar
feats
ex
hib
it
ed
via
trans
fer
o
r
act
ivati
on fu
nction as i
n (
5
)
bel
ow:
∅
=
(
)
=
∑
∗
=
1
(5)
ANN
at
te
mp
ts
to
tran
sla
te
into
m
at
hemati
cal
model,
pri
nciples
of
bio
l
og
ic
al
processi
ng
so
as
to
gen
e
rate
i
n
t
he
fastest
ti
me,
i
mp
li
ci
t
predict
ive
ou
tc
om
es
of
a
ta
s
k
[
42
-
43].
Its
outc
ome
s
a
re
der
i
ve
d
from
exp
e
rience
,
an
d
it
is
able
to
re
cognize
feats
a
nd
be
hav
i
ours
of
i
nterest
f
rom
histo
ric
data
se
t
–
to
yield
an
opti
mal
so
luti
on
of
high
qual
it
y
a
nd
vo
i
d
of
over
-
fi
tt
ing
,
irres
pect
ive
of
m
od
i
ficat
ion
via
oth
e
r
appr
ox
imat
io
ns
with
mu
lt
iple
a
ge
nts.
T
hese
al
s
o,
const
antly
af
fe
ct
s
the
qual
it
y
of
a
ny
s
olu
ti
on
[
44].
Its
c
onfi
gurati
on
de
pends
on
t
he
area
to
be
app
li
ed
,
capt
ured
data
feats
a
nd
s
ys
te
m
requireme
nt.
Its
c
onnecti
ons
are
set
as
ei
ther
exp
li
ci
t
(aprio
ri
knowle
dg
e
)
and/o
r i
mp
li
ci
t (
post
-
pri
or
i
knowle
dg
e)
to
all
ow
le
ar
ning
so t
hat
the
net is
trai
ne
d t
o l
earn
patte
rn
s
that
ch
ang
e
it
s w
ei
ght
an
d
bias
base
d
on
a
r
ule [40].
Its
le
a
rn
i
ng
is gro
up
e
d
i
nto
ei
ther
of:
s
up
e
r
vised
,
uns
uper
vise
d
a
nd r
ei
nfor
ce
me
nt [4
5,
46].
The
nat
ur
e
of
mar
ket
data
is
chao
ti
c
an
d
re
qu
i
res
previ
ous
knowle
dge.
Th
us
,
we
ad
opt
the
recurre
nt
(
Jor
dan
)
netw
ork
so
t
hat
it
i
nc
or
porates
pre
vious
dataset
a
nd
pr
e
vious
outpu
t
to
be
fee
dback
as
i
nput
i
nto
the
model’s
hidde
n
un
it
s,
as
in
put
into
m
odel
[
47,
48]
t
o
yield
nex
t
outp
ut.
Its
co
rr
el
at
e
d
weig
hts
(
W
i.j
)
betwee
n
the
in
put
a
nd
hi
dd
e
n
la
ye
rs
,
bias
(W
o
j)
a
nd
the
ma
r
ket
bas
ket
anal
y
s
is
dataset
(
x
i
)
is
s
ummed
vi
a
the
ta
ng
e
nt/si
gmoid tra
nsfer
f
un
ct
i
on
t
o yiel
d
it
s
ou
t
pu
t a
s in
(
6
)
and
(
7
)
[49].
=
+
∑
∗
=
1
(6)
(
)
=
2
1
+
−
2
∗
−
1
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Inf & C
ommu
n Tec
hn
ol
IS
S
N:
22
52
-
8776
In
ve
nto
ry
pred
ic
ti
on
an
d man
ag
e
me
nt in
Ni
ge
ria
us
in
g mar
ket
b
ask
et
an
alysis …
(
Ar
nold
Ad
i
mab
ua O
j
ugo
)
131
We
c
on
st
ru
ct
our
J
orda
n
net
by
m
od
i
fy
i
ng
t
he
mu
lt
il
ayer
e
d
feedfo
r
ward
wi
th
ad
diti
on
a
c
ont
ext
la
yer
to
he
lp
retai
n
data
betwee
n
ob
s
er
vations.
With
eac
h
move,
ne
w
i
nputs
are
fe
d
i
n
a
nd
pr
e
vious
co
nt
ents
i
n
hidden
la
ye
r
is
passe
d
i
nto
c
onte
xt
la
ye
r,
an
d
la
te
r
fed
bac
k
int
o
the
hi
dden
la
ye
r
i
n
the
ne
xt
ti
me
-
ste
p.
Th
e
con
te
xt
la
ye
r
a
t
sta
rt
is
init
ia
l
iz
ed
to
ze
ro
–
so
that
outp
ut
from
t
he
hidde
n
la
ye
r
on
the
first
it
erati
on
will
be
same as i
f
the
r
e is n
o
c
onte
xt
la
yer
[50
].
The
net
res
ol
ve
s
str
uctu
ral
de
pende
ncies
i
mpose
d
on
it
by
dataset
a
nd
hybri
d
he
uri
sti
cs
us
e
d
via
it
s
abili
ty
to
st
or
e
earli
er
dat
a
a
s
ge
ner
at
e
d
from
previ
ou
s
la
yer(s)
[
51].
F
eed
-
for
ward
ne
ts
are
e
xpan
de
d
a
nd
exten
ded
t
o
r
epr
ese
nt
co
mpl
ex
dyna
mic
patte
rn
s
(as
our
data
is
ri
pple
d
with
ne
w
a
nd
previ
ous
set
s).
Feed
forw
a
r
d
n
et
s
treat
al
l
dat
a
as
new
s
o
tha
t
pr
e
vious
data
set
can
no
t
hel
p
t
he
model
i
de
ntify
data
feats
,
eve
n
if
s
uc
h
datase
ts
ex
hib
it
s
te
mporal
de
pe
ndence;
ca
us
in
g
pract
ic
al
di
f
ficult
y
as
net
work
bec
om
es
la
r
ger.
Howe
ver,
Jor
dan
netw
ork
overcomes
this
di
ff
ic
ulty
thr
ough
it
s
inter
nal
f
eedb
ac
ks
–
ma
king
it
ap
pro
pri
at
el
y
su
it
able
fo
r
dy
namic,
no
n
-
li
ne
ar
a
nd
c
omplex
ta
s
ks
.
T
hu
s
,
ou
t
pu
t
is
fe
d
ba
ck
as
in
put
int
o
hidde
n
la
ye
r
with
a
ti
me d
el
ay
[52
-
53].
Our
rati
onal
e
f
or
the
Jo
r
dan’s
net
wor
k
is
be
cause
it
is
m
or
e
plausib
le
a
nd
c
ompu
ta
ti
onal
ly
mor
e
powe
rful
tha
n
oth
e
rs
due
t
o
use
of
bac
kp
r
opagati
on
-
in
-
ti
m
e
le
ar
ning
s
o
that
it
s
ou
t
put
a
t
ti
me
t
is
us
e
d
al
ong
with
ne
w
i
nput
data
to
c
omp
ute
it
s
ou
t
pu
t
at
ti
me
t+
1
in
res
ponse
t
o
model’
s
dynamic
an
d
non
-
li
near
feat
s
[
51]
.
Ou
t
pu
t
is
c
ompu
te
d
via
Ta
nsi
g
f
unct
io
n
y
k
,
w
hich
s
um
s
in
pu
t,
recei
ve
s
ta
r
get
val
ue
of
trai
ning
pa
tt
ern
,
com
pu
te
s
er
ror
data
as
well
as
updates
weig
ht
c
j
k
an
d
bias
c
o
k
.
Er
ror
is
se
nt
bac
k
i
n
ne
xt
move
t
o
in
pu
t
nodes
from
outp
ut
vi
a
er
ror
-
bac
kpr
op
a
gatio
n
to
c
orrect
t
he
wei
gh
ts
a
nd
fi
nd
t
ho
s
e
t
hat
a
ppr
ox
imat
es
to
th
e
ta
r
get
ou
t
pu
t
with
sel
ect
ed
accu
rac
y.
Weig
hts
are
m
od
i
fied
by
mini
mizi
ng
e
rror
be
tween
ta
r
get
a
nd
c
ompu
te
d
out
pu
ts
as
forwar
d
pa
s
s
en
ds
.
I
f
the
error
is
higher
than
sel
ect
ed
value,
proces
s
con
ti
nues
with
re
verse
pas
s
;
el
se,
trai
ning sto
ps [
54
-
55].
Its
trai
ning
ai
m
at
best
fit
w
ei
gh
t
dataset
t
hat
ass
um
e
s
a
ppr
oximat
ion
in
flue
nce
of
data
points
at
t
he
center
–
s
o
t
ha
t
f
un
ct
io
n
de
creases
with
di
sta
nce
f
r
om
it
s
cente
r.
Its
E
uclidean
le
ng
t
h
(r
j
)
yields
distance
betwee
n datu
m
v
ect
or
y
=
(y
1
,
...
,
y
m
)
and ce
nt
er (w
1j
,...,
w
mj
)
as
in
(
8
)
[48,
49]:
=
|
|
−
|
|
=
{
∑
(
−
)
2
=
1
}
½
(8)
The
s
uitable
tr
ansf
e
r funct
io
n i
s appli
ed
t
o r
j
to
yield
(
9
)
:
∅
(
)
=
∅
|
|
−
|
|
(9)
Finall
y,
outp
ut
k recei
ves
wei
gh
te
d
c
ombina
ti
on
as
in (
10
)
:
=
+
∑
(
∗
∅
(
)
)
=
=
1
+
∑
(
∗
∅
|
|
−
|
|
)
=
1
(10)
2.2.
Gene
tic
a
lgo
ri
t
hm
(G
A)
G
A
is
ins
pire
d
by
Da
r
winian
ge
netic
ev
olu
ti
on
(sur
viv
al
of
fi
tt
est
)
co
ns
ist
s
of
popula
ti
on
(
da
ta
)
ch
os
e
n
for
sel
ect
ion
with
pote
ntial
so
luti
ons
to
a
sp
eci
fic
ta
sk.
Each
po
te
ntial
so
luti
on
is
an
ind
ivi
du
al
for
wh
i
c
h
op
ti
mal
is
f
ound
usi
ng
f
our
op
erators:
init
ia
lize,
sel
ect
,
cr
ossove
r
a
nd
m
utati
on
[
33,
56].
I
nd
i
viduals
wit
h
ge
nes
cl
os
e
to
opti
m
al
,
is
sai
d
t
o
be
fit.
Fit
ne
ss
f
un
ct
io
n
deter
m
ines
how
cl
os
e
an
i
nd
i
vidual
is
to
op
ti
mal
s
olu
ti
on.
Ojugo,
A.A
,
et a
l
[38] no
te
s t
he
operat
or
s
as:
a.
In
it
ia
li
ze
–
Ind
ividu
al
data
a
r
e
e
ncode
d
i
nto
f
orms
s
uitable
f
or
sel
ect
io
n.
Each
enc
odin
gs
ty
pe
use
d
ha
s
it
s
merit
.
Bi
nary
e
nc
od
i
ngs
a
r
e
co
mputat
io
na
ll
y
m
or
e
e
xp
e
nsi
ve.
Decimal
en
co
ding
ha
s great
er d
i
ver
sit
y
in
c
hrom
osom
e
a
nd
gr
eat
er
va
riance
of
po
ols
gen
e
rated;
fl
oat
-
po
i
nt
e
nc
odin
g
or
it
s
c
ombinati
on
is
mor
e
eff
ic
ie
nt
t
han
bin
a
ry.
T
hus,
it
enc
od
es
as
fi
xed
le
ngth
vec
tors
f
or
one
or
m
or
e
po
ols
of
dif
fer
e
nt
t
yp
e
s.
The
fi
tnes
s
fun
ct
ion
e
valuates
ho
w
cl
ose
a
s
olu
ti
on
is
to
it
s
opti
mal
–
a
fter
wh
ic
h
t
hey
are
c
hosen
f
or
reprod
uction.
I
f
s
olu
ti
on
is
found,
functi
on
i
s
go
od
;
el
se,
is
bad
an
d
not
sel
ect
ed
for
c
r
os
s
ov
e
r.
T
he
fitne
ss
functi
on is th
e
on
l
y part
with
knowle
dge
of task.
If m
or
e
s
ol
ution
s
are f
ound, th
e
h
i
gh
e
r
i
ts fit
ness value
.
b.
Sele
ct
ion
–
bes
t
fit
ind
ivid
uals
cl
os
e
to
op
ti
m
al
are
ch
os
e
n
to
mate
.
T
he
la
r
ger
t
he
num
be
r
of
sel
ect
e
d,
th
e
bette
r
t
he
c
ha
nc
es
of
yieldi
ng
fitt
er
in
div
i
du
a
ls.
T
his
c
o
ntinues
un
ti
l
on
e
is
c
ho
s
en
,
from
th
e
la
st
tw
o/
th
ree
remaini
ng
so
l
ut
ion
s,
t
o
becom
e
sel
ect
ed
pa
re
nts
to
new
offs
pr
i
ng.
Sele
ct
io
n
e
ns
ures
th
e
fi
tt
est
ind
ivid
ual
s
are
c
hosen
f
or
mati
ng
but
al
s
o
al
lo
ws
f
or
le
ss
fit
i
nd
i
vidua
ls
from
the
pool
a
nd
the
fitt
est
to
be
sel
ect
ed
.
A
sel
ect
io
n
tha
t on
l
y
mate
s
th
e fitt
est
is
el
it
is
t
and ofte
n l
eads t
o
c
onvergi
ng at local
opti
ma.
c.
Cros
s
over
e
nsures
best
fit
in
div
id
ual
ge
nes
are
e
xch
a
nge
d
t
o
yield
a
ne
w,
fitt
er
po
ol.
T
her
e
a
re
t
w
o
cro
ss
over
ty
pe
s
(
dep
e
nds
on
encodin
g
t
yp
e
us
e
d)
:
(a)
sim
pl
e
cro
s
sover
f
or
bi
nary
e
nc
oded
pool.
I
t
al
lo
ws
sing
le
-
or
m
ulti
-
point
c
ross
w
it
h al
l
ge
nes
f
r
om
a
par
e
nt,
a
nd
(
b)
ar
it
hmet
ic
crosso
ve
r a
ll
ow
s
new
pool
t
o
be
c
reated
by a
dd
i
ng an i
ndivi
du
al
’s perce
nta
ge
to
anothe
r.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
In
t J
Inf & C
ommu
n Tec
hn
ol
,
V
ol.
8
,
No.
3,
Dec
201
9
:
12
8
–
138
132
d.
M
utati
on
al
te
r
s
ch
romo
some
s
by
c
ha
ng
i
ng
it
s
gen
e
s
or
it
s
seq
ue
nce,
t
o
ens
ur
e
ne
w
po
ol
c
onverges
to
global
mi
nima
(instead
of
loc
al
opti
ma).
Algorithm
sto
ps
i
f
opti
mal
is
found,
or
afte
r
num
ber
of
r
uns
if
new
po
ols
are
create
d
(th
ough
c
ompu
ta
ti
on
al
ly
ex
pensi
ve
),
or
w
hen
no
bette
r
s
olu
ti
on
is
fou
nd.
G
ene
s
may
c
ha
nge
ba
sed
on
prob
a
bili
ty
of
m
utati
on
rate.
Muta
ti
on
im
pro
ves
the
m
uch
needed
div
e
rsity
i
n
reprod
uction a
nd it
s alg
or
it
hm i
s as t
hus:
Cult
ur
al
GA
is
a
var
ia
nts
of
GA
with
a
belie
f
sp
ac
e
de
fine
as
thus:
(a)
N
ormat
ive
(
has
s
pecific
val
ue
ranges
to
w
hich
an
i
nd
i
vidual
is
bound),
(
b)
Domai
n
(h
as
da
ta
about
ta
sk
domain
),
(c)
Te
mporal
(
has
dat
a
about
even
ts
’
sp
ace
i
s
a
vaila
ble),
a
nd
(
d)
S
patia
l
(
ha
s
to
pogra
ph
ic
al
data)
.
I
n
a
dd
it
ion
,
an
in
flue
nce
f
un
ct
io
n
m
ediat
es
betwee
n
belie
f
sp
ace
a
nd
the
pool
–
to
en
sure
and
al
te
r
in
div
i
du
al
s
i
n
the
po
ol
to
c
onf
orm
to
belie
f
spa
ce.
CGA
is
cho
se
n
t
o
yi
el
d
a
po
ol
that
do
e
s
not
vi
olate
it
s
belie
f
sp
a
ce
and
hel
ps
re
du
ce n
umber
of
possi
ble
ind
i
vidual
s
GA g
e
ne
rates ti
ll
an
opti
mum
is fou
nd [5
6,
34, 3
7,
38]
.
3.
MA
TE
RIA
L
S
AND MET
H
ODO
L
OG
Y
3.1.
Problem
desc
ri
pt
ion
a
nd fo
rmul
at
i
on
Con
si
der
mar
ke
t
data
logs
tha
t
include i
te
ms
purc
hased
by
c
us
to
mers
. T
he
mana
ger
of
a
s
up
e
rma
rk
et
wan
ts
t
o
m
axi
mize
the
int
ere
sti
ng
ness
of
th
e
pro
duct
place
ment
on
s
helve
s.
T
he
inte
resti
ngness
val
ue
is
relat
ed
to
mi
ned
ass
oc
ia
ti
on
ru
le
s
a
nd
eac
h
it
em
’s
locat
io
n
on
s
helves
[
57].
T
he
rati
onal
e
f
or
the
intere
sti
ngne
ss
maximiza
ti
on
with
locat
io
n
c
on
si
der
at
io
ns
i
s
base
d
on
the
fact
that,
ass
oc
ia
ti
on
r
ule
min
ing
helps
maxi
mize
cro
ss
-
sel
li
ng
ef
fect
of
it
ems
[
58].
It
is
al
so
cl
ear
that
the
lo
cat
ion
o
f
s
helv
es
has
the
unde
niable
impact
on
the
sel
li
ng
rate
of
it
ems.
T
hus,
it
ems
that
are
pl
aced
near
the
entra
nce
or
e
xit
do
or
s
hav
e
m
or
e
cha
nce
to
be
purc
hased.
S
o,
pr
e
fer
e
nce
f
un
ct
ion
of
the
sto
re’
s
ma
na
ger
de
pends
on
t
he
f
ollow
i
ng
par
a
mete
rs:
sel
li
ng
ben
e
fit,
su
pp
or
t
a
nd
c
onfi
den
ce
of
ea
ch
pair
of
it
em
s,
an
d
the
sel
li
ng
possibil
it
y
of
eac
h
it
em
from
eac
h
sh
el
f
[59]
.
These
pa
ramet
ers
a
re thus i
ntegr
at
e
d
int
o
th
e pre
fe
ren
ce
fu
nction (
pf)
as
in (
11
)
:
=
∑
[
∑
[
+
∑
[
+
]
=
1
]
=
+
1
]
−
1
=
1
(11)
m
is
t
he
num
be
r
of
it
ems,
p
is
the
numb
e
r
of
sh
el
ves
,
C
il
is
the
c
onfide
nce
of
the
r
ule
(ite
m
i
→
it
em
l
),
b
i
is
sel
li
ng
ben
e
fit
of
t
he
i
th
it
em,
v
ik
is
t
he
sel
li
ng
po
s
s
ibil
it
y
de
gr
ee
of
the
it
em
i
if
a
nd
w
he
n
place
d
int
o
the
k
th
-
s
helf,
a
nd
x
ik
is
bin
a
ry
decisi
on
va
ria
ble
that
ta
kes
va
lue
of
1
w
he
n
the
it
em
i
is
al
locat
ed
to
t
he
s
helf
k
;
Othe
rw
ise
,
it
s
value
is
0.
T
he
re
are
restrict
io
ns
that
li
mit
pref
ere
nce
f
unct
ion
value
.
T
hu
s
,
capaci
ty
li
mit
at
ion
(
cl
) of ea
ch
sh
e
lf m
us
t be
cons
idere
d
as t
he fo
ll
ow
in
g
c
onstr
ai
nt:
=
∑
≤
ℎ
=
1
,
2
,
…
=
1
(12)
U
k
is
the
ca
pacit
y
of
the
k
th
-
s
he
lf.
T
he
sec
ond
co
ns
trai
nt
is
th
e
ass
ociat
ion
c
on
st
raint
su
c
h
that
s
uppor
t
of
the
ru
le
(ite
m
i
→
it
em
l
)
mu
st
be
great
er
tha
n
minim
um
th
reshold
de
te
rmin
e
d
by
th
e
decisi
on
mak
er.
T
he
ob
je
ct
ive
f
un
ct
ion
a
nd
c
onstr
ai
nts
are
non
-
l
inear
functi
ons
in
wh
ic
h
decisi
on
va
riables
are
bin
ar
y.
T
hus,
we
will
d
eal
w
it
h
a rou
gh
feasi
bl
e sp
ace t
hat inc
reases t
he
pr
obabili
ty of tra
pping
i
n
the
local
opti
mu
m
. T
hu
s,
ou
r
need f
or the
use
of a
n
e
vo
l
ution
a
r
y un
s
uper
vised
m
od
el
i
n t
he
sce
nar
io
th
us
pr
es
ente
d.
3.2.
Nu
meri
c
ex
am
ple d
ataset
Dataset
is
retri
eved
from
Del
ta
M
al
l
(Shop
r
it
e)
Asa
ba
an
d
Wa
rr
i
res
pecti
vely
dataset
as
in
T
a
ble
1
belo
w.
Ta
ble
1
shows
t
he
e
nc
od
e
d
m
ar
ket
ba
sk
et
dataset
of
it
ems
as
they
a
re
co
-
sel
ect
ed
of
t
he
va
rio
us
s
helves
and
placed
i
n
the
bas
ket
at
th
e
same
ti
me.
F
or
e
xam
ple,
S
01
f
or
Item
1
ha
s
a
pr
e
valence
of
0.81.
Th
is
i
mp
li
es
that
there
is
81%
cha
nce
that
i
te
ms
1,
2,
6 a
nd
8 a
re
picke
d
from s
helf
S01
and
place
d i
n t
he ba
sk
et
.
The
Delt
a
M
al
l
ma
r
ket
ba
sk
et
dataset
was
em
ployed
to
sim
ulate
t
he
m
od
el
as
well
as
yield
c
um
descr
i
be
t
he
pr
opos
e
d
model
-
based
s
olu
ti
on.
T
hu
s
,
t
he
s
ys
te
m
s
hows
ei
gh
t
it
ems
that
m
us
t
be
al
l
ocated
into
f
ou
r
-
s
helves
.
Also,
based
on
the
sh
el
f’
s
posit
ion
s
,
eac
h
s
helf
has
a
diffe
ren
t
im
pact
on
the
sel
li
ng
po
s
s
ibil
it
y
of
al
loca
te
d
goods,
a
nd
these
po
s
sibil
it
ie
s were
determi
ned by ec
onom
ist
s
and e
xp
e
rts as
presente
d
in
T
able 1.
Table
1.
E
nc
od
ed
bin
a
ry r
e
pr
e
sentat
ion
of
5
-
bas
kets/shelve
values
for
anal
ys
is
Item
s
1
2
3
4
5
6
7
8
Sh
elv
es
S0
1
1
1
0
0
0
1
0
1
S0
2
0
1
0
1
1
0
1
0
S0
3
0
0
1
0
1
0
1
1
S0
4
1
0
1
1
1
0
0
0
(Sou
rce:
Au
th
o
rs’
o
wn
pro
cess
in
g
an
d
tr
an
slatio
n
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Inf & C
ommu
n Tec
hn
ol
IS
S
N:
22
52
-
8776
In
ve
nto
ry
pred
ic
ti
on
an
d man
ag
e
me
nt in
Ni
ge
ria
us
in
g mar
ket
b
ask
et
an
alysis …
(
Ar
nold
Ad
i
mab
ua O
j
ugo
)
133
3.3.
Model
desi
gn
p
roblems
Issu
e
s to
b
e
r
es
olv
e
d
in
the
m
od
el
desi
gn inc
lud
e:
a.
M
a
ny
st
ud
ie
s
a
im
at
sing
le
he
ur
ist
ic
to
gl
ob
a
ll
y
cl
assify
dat
a
or
r
ules
int
o
var
i
ou
s
cl
asses
.
This
ha
s
oft
e
n
yielded
false
-
posit
ives
(classi
fy
i
ng
a
ru
le
as
genuine
w
hen
it
is
false)
an
d
tru
e
-
neg
at
iv
es
(in
a
bili
ty
of
model t
o
cl
as
sify a r
ule)
er
ror result
s.
b.
Su
c
h
models
e
mp
lo
y
hill
cl
im
bing
meth
ods
t
hat
of
te
n
gets
t
heir
so
l
ution
tr
app
e
d
at
lo
cal
minima
beca
use
their s
peed s
hr
i
nk
s
as
s
uc
h mo
dels
of
te
n ap
pr
oach
e
s it
s opti
ma.
c.
Re
so
lvin
g
co
nfl
ic
t
issues
in
str
uctu
red
le
ar
ning
a
nd
f
r
om
sta
ti
sti
cal
dep
en
de
ncies
im
posed
by
data
an
d
t
he
us
e
of m
ulti
ple meth
ods a
dopt
ed/ada
pted
, is
qu
it
e a te
dious.
3.4.
Model
desi
gn
goals
an
d
obje
ctive
The
pro
pose
d
sy
ste
m
ai
ms
to
so
l
ve
t
he
e
xi
sti
ng
pro
blem
of
ma
rk
et
ba
sk
et
a
nalysis
util
iz
ing
t
he
fo
ll
owin
g pro
pe
rtie
s,
w
hich
is
in
ta
ndem
to
[60
-
62]:
a.
Embo
dy
the
kn
ow
le
dg
e
of
hu
man
ex
pe
rts
with
the
help
of
s
pecial
s
of
twa
re
to
ols,
mani
pu
l
at
e
data
t
o
s
olve
pro
blems a
nd
make deci
sio
ns i
n
that
domai
n.
b.
Pr
oc
esses a
re
be
tt
er f
ormal
iz
e
d
a
nd d
e
fin
e
d on mac
hin
es
.
c.
Kno
wled
geb
a
s
e upd
at
e is
au
t
om
at
ic
d.
Pr
oc
esses a
re
be
tt
er f
ormal
iz
e
d
a
nd d
e
fine
d on mac
hin
es
.
3.5.
Ex
peri
menta
l
mod
el
/
algori
th
m
fr
amewo
rk
The pr
opose
d
model co
ns
ist
s
of fo
ur p
a
rts:
a.
Kno
wled
geb
a
s
e
co
ns
ist
s
of
hi
storic,
obser
ve
d
-
st
ru
ct
ur
e
d
i
te
ms
co
-
occ
urr
ence
dataset
(
f
eat
s)
of
ma
r
ket
bas
ket
for
Del
ta
M
al
l.
These
hav
e
bee
n
gr
acefully
e
nc
oded
as
if
-
t
hen
r
ules
–
her
e
by
represe
nted
as
op
ti
mize
d bina
ry fun
ct
io
ns f
or the
selec
te
d d
at
a feats.
b.
Infer
e
nce
e
ng
i
ne
c
on
sist
s
of
hybri
d
ass
ociat
ive
r
ules
an
d
t
he
gen
et
ic
al
gorith
m
trai
ne
d
neural
net
wor
k
model.
T
hus,
t
he
in
fer
e
nce
e
ng
i
ne
see
ks
t
o
infe
r
co
ns
e
qu
ents
de
rive
d
from
a
ntecede
nt
s
that
ha
ve
be
en
trai
ned
us
i
ng
the
hybr
i
d
m
em
et
ic
al
go
rith
m.
The
ru
le
s
re
pr
e
sent
sel
ect
ed
da
ta
feats
of
i
nterest
enc
oded
as
if
-
the
n
(
r
ule
-
ba
sed)
co
ndit
ion
s
with
po
s
sib
le
ou
tc
om
es
a
nd
act
ion
s
(cla
ssifie
d
int
o
s
uppo
rt,
co
nfi
de
nce
and
li
ft
cl
asses
)
upon
crit
eri
on score
bei
ng met or
ac
hieve
d.
T
he
Jor
dan n
et
w
ork
pr
ov
i
de
s
a
sel
f
-
le
ar
nin
g
machine
,
bette
r
t
un
e
d
f
or
r
obus
tness
via
ge
ne
ti
c
al
g
ori
thm
op
ti
mize
r
t
hat
yields
great
er
f
le
xib
il
it
y
of
th
e
ru
le
-
base
d
dat
a.
T
hus,
it
a
da
pts
the
s
ys
te
m
to
aut
onom
ou
sl
y
cl
assi
fy
the
mar
ket
ba
sk
et
data
int
o
varyin
g
cl
ass
-
typ
e
s
as
well
as
yield
ce
ntral
iz
ed
-
scal
e
d
bo
unda
ry
i
n
determini
ng
hi
gh
or
l
ow
de
gr
e
e
mem
be
rs
hip
f
unct
ion.
c.
Decisi
on
s
upport
–
co
ns
ist
s of
the
p
re
dicte
d ou
t
pu
t
an
d
the
o
ut
pu
t
data
bas
e
that
is
up
dated
a
utomat
ic
al
ly
in
ti
me
as
patie
nts
are
dia
gnos
es
as
lo
ng
a
s
it
enc
ounters
an
d
r
ead
sin
new
data.
T
he
decisi
on
s
uppo
rt
pr
e
dicts
s
ys
te
m
ou
t
put
base
d
on
the
co
gnit
ive
a
nd
the
e
mo
ti
onal
file
rs
as
dis
play
by
the
ou
tp
ut
dev
i
ce.
This is see
n
i
n Fi
gure
2.
M
odel
is
fi
rst
init
ia
li
zed
with
ru
le
s.
I
nd
i
vidu
al
so
luti
ons
ar
e
sel
ect
ed
f
rom
pool
via
t
ou
r
na
men
t
m
et
hod
to
dete
rmin
e
t
he
ca
nd
i
dates
to
mate
a
nd
yield
ne
xt
ge
nerat
ion
.
C
r
os
sov
er
a
nd
m
utati
on
is
a
ppli
ed
t
o
help
netw
ork
le
ar
n
dynamic
a
nd
non
-
li
near
feats
in
the
dataset
a
nd
feats
of
inte
rest
us
i
ng
a
m
ulti
-
point
cr
os
s
ov
e
r.
With
m
utati
on
,
da
ta
a
re
ra
ndoml
y
ge
ner
at
e
d
us
in
g
Ga
us
sia
n
dist
rib
ution
c
orres
pondin
g
to
cro
ss
over
point
s
(all
gen
e
s
are
from
sing
le
par
e
nt).
As
new
par
e
nts
con
t
rib
utes
to
yield
ne
w
pool,
mu
ta
ti
on
is
ap
plied
to
yield
ra
ndom
gen
e
s
from
w
hi
ch
three
-
can
di
dates
are
sel
ect
ed
(a
nd
al
locat
ed
ne
w
ra
ndom
values
t
hat
conf
or
m
s
to
bo
unda
ry
li
mit
s)
to un
de
rgo furt
he
r
m
ut
at
ion
. T
he
num
ber
of m
utati
on a
pp
li
ed
d
e
pe
nd
s
on h
ow f
a
r
GA is
progres
sed on
the
netw
ork
(how
fit
is
t
he
fitt
est
ind
ivi
du
al
in
the
po
ol),
w
hich
eq
uals
fitness
of
the
fitt
est
ind
i
vidual
div
ide
d
by
2.
New
i
nd
i
viduals
re
place
old
with
l
ow
fitness
so
as
t
o
cr
eat
e
a
new
po
ol.
Process
c
on
ti
nu
e
s
unti
l
ind
ivid
ua
l
with
a
fitness
of
0
is
fou
nd.
Th
us
,
s
olu
ti
on
has
bee
n
rea
ch
ed.
I
niti
al
iz
at
ion
/Sel
ect
ion
with
A
NN
e
nsure
s
the
first
3
-
belie
fs
are
met;
w
hile,
m
utati
on
en
s
ur
es
f
ourth
-
belie
f
is
met.
Its
influ
e
nce
f
un
c
ti
on
determi
ne
s
the
numb
e
r
of
m
utati
on
s
that
ta
ke
place
,
a
nd
kn
ow
le
dg
e
of
s
ol
ution
(i.e.
how
cl
os
e
s
olu
ti
on
is)
has
direct
i
mp
act
on ho
w
m
odel
is processe
d. T
he GA
N
N
m
od
el
p
se
udoc
od
e
is as th
u
s:
IN
P
UT:
1.
Pool siz
e (
k), c
ro
ss
over
(
c
),
m
utati
on
(
v)
,
in
fluen
ce
fu
nctn (I
fn
c
)
a
nd n
;
//
I
niti
alizati
on an
d Selec
ti
on
2.
Ra
ndom
l
y gene
rate
K
possi
ble so
l
ution
3.
Save sol
ution i
n pop
ulati
on
K
ok;
//
Loop ti
ll
terminal
po
i
nt
4.
Fo
r
m
=
1
to
n
do
;
//
Cr
os
s
ove
r
5.
Numbe
r of cr
osso
ver
nc
= (k
–
Ifnc)
/
2;
6.
Fo
r
u =
1
t
o n do;
7.
Sele
ct
two
so
l
ut
ion
s
rand
om
ly
E
A
an
d F
G
f
or
K;
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
In
t J
Inf & C
ommu
n Tec
hn
ol
,
V
ol.
8
,
No.
3,
Dec
201
9
:
12
8
–
138
134
8.
Gen
e
rate
G
V
a
nd H
N
by 2
-
po
i
nt cross
over t
o E
A
an
d
F
G;
9.
Save G
V
a
nd H
N
to
K2;
10.
En
d
F
or
;
//
M
uta
ti
on
11.
Fo
r
u =
1
t
o n do;
12.
Sele
ct
ion
a
so
l
ution Y
h
fr
om
K2
;
13.
M
utate
eac
h bi
t of Y
h
under I
f
nc
14.
Ge
ner
at
e a
n
e
w
s
olu
ti
on
Y
h
i
15.
If
Y
h
i
is i
mpos
sible
16.
Re
com
pute
Y
h
i
with
po
ssi
ble
so
luti
on
by m
odif
ying
Y
h
i
17.
End if
18.
Re
com
pute
Y
h
with
Y
h
i
in
K2
19.
En
d
f
or
//
Reco
mpute
20.
Re
com
pu
te
K
=
K2
;
21.
R
et
ur
n
Best
s
ol
ution
in
Y
M
odel
sto
ps
i
f
stop
c
rite
rio
n
z
ero
(0)
is
met;
Or
,
if
numb
e
r
of
e
pochs
is
re
ached.
T
hus,
num
ber
of
set
s
of size
k
picke
d from
p
it
ems
yields:
(
)
=
!
!
(
−
)
!
≡
8
!
4
!
(
8
−
4
)
!
=
105
3.6.
Result
findin
gs an
d d
isc
ussi
on
The
res
ult
s
im
ulate
d
ass
ociat
ion
of
ba
sk
et
a
nalysis
dataset
is
sho
wn
in
T
able
2.
Anothe
r
featu
re
o
f
major
c
on
c
er
n
and
im
pact
t
o
the
mar
ket
bas
ke
t
data
in
the
a
ll
ocati
on
of
it
ems
t
o
sh
el
ves
is
the
sel
li
ng
be
nef
i
t
.
So
,
it
is
lo
gical
that
i
n
ma
xim
iz
ing
e
xpect
ed
ben
e
fit
of
the
s
el
li
ng
,
t
he
pro
duct
s
with
the
hi
gh
e
r
be
nef
it
s
mu
st
be
al
locat
ed
to
sh
el
ves
with
hi
gh
e
r
sel
li
ng
po
ssibil
it
ie
s.
Table
3
s
hows
t
he
va
lues
of
the
pro
du
ct
s
’
ben
e
fit.
Table
4
a
nd Ta
ble 5 s
how
t
he
s
uppo
rt and c
onfide
nc
e simulat
ed
val
ues
re
sp
ect
iv
el
y.
Table
2.
Simul
at
ed
ass
ociat
io
n of bas
ket a
na
lysis dataset
Item
s
1
2
3
4
5
6
7
8
Sh
elv
es
S0
1
0
.81
0
.43
0
.12
0
.90
0
.22
0
.25
0
.34
0
.87
S0
2
0
.51
0
.21
0
.55
0
.19
0
.90
0
.71
0
.12
0
.21
S0
3
0
.32
0
.82
0
.61
0
.28
0
.21
0
.23
0
.19
0
.12
S0
4
0
.54
0
.76
0
.11
0
.58
0
.31
0
.46
0
.01
0
.40
(Sou
rce:
Au
th
o
rs’
o
wn
pro
cess
in
g
an
d
tr
an
slatio
n
)
Table
3.
Sell
ing
po
s
sibil
it
y
of
each
it
em
to p
ur
c
hase
d (V
ik
:
i
= 1,2
,…
8; k =
1
,
2,3,4
)
Item
s
1
2
3
4
5
6
7
8
Sh
elv
es
S0
1
0
.81
0
.43
0
.12
0
.90
0
.22
0
.25
0
.34
0
.87
S0
2
0
.51
0
.21
0
.55
0
.19
0
.90
0
.71
0
.12
0
.21
S0
3
0
.32
0
.82
0
.61
0
.28
0
.21
0
.23
0
.19
0
.12
S0
4
0
.54
0
.76
0
.11
0
.58
0
.31
0
.46
0
.01
0
.40
(Sou
rce:
Au
th
o
r’
s
o
wn
pro
cess
in
g
)
Table
4.
Sup
port v
al
ues of si
mu
la
te
d data
(
S
il
)
Item
s
1
2
3
4
5
6
7
8
Sh
elv
es
01
0
.42
0
.31
0
.12
0
.09
0
.14
0
.28
0
.13
0
.42
02
0
.00
0
.39
0
.18
0
.11
0
.19
0
.34
0
.19
0
.29
03
0
.00
0
.00
0
.46
0
.19
0
.22
0
.11
0
.35
0
.32
04
0
.00
0
.00
0
.00
0
.41
0
.18
0
.07
0
.19
0
.21
05
0
.00
0
.00
0
.00
0
.00
0
.36
0
.02
0
.21
0
.01
06
0
.00
0
.00
0
.00
0
.00
0
.00
0
.49
0
.21
0
.12
07
0
.00
0
.00
0
.00
0
.00
0
.00
0
.00
0
.44
0
.32
08
0
.00
0
.00
0
.00
0
.00
0
.00
0
.00
0
.00
0
.32
(So
u
rce
:
A
u
t
h
o
r’
s
o
w
n
pr
o
ce
s
s
i
n
g
)
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Inf & C
ommu
n Tec
hn
ol
IS
S
N:
22
52
-
8776
In
ve
nto
ry
pred
ic
ti
on
an
d man
ag
e
me
nt in
Ni
ge
ria
us
in
g mar
ket
b
ask
et
an
alysis …
(
Ar
nold
Ad
i
mab
ua O
j
ugo
)
135
Table
5.
C
onfi
den
ce
v
al
ues o
f
sim
ulate
d dat
a (C
il
)
Item
s
1
2
3
4
5
6
7
8
Sh
elv
es
01
1
0
.52
0
.43
0
.24
0
.35
0
.49
0
.13
0
.10
02
0
.52
1
0
.35
0
.25
0
.63
0
.03
0
.21
0
.30
03
0
.43
0
.33
1
0
.20
0
.38
0
.39
0
.13
0
.21
04
0
.23
0
.24
0
.20
1
0
.67
0
.32
0
.13
0
.12
05
0
.36
0
.65
0
.39
0
.67
1
0
.88
0
.31
0
.26
06
0
.49
0
.01
0
.41
0
.32
0
.87
1
0
.11
0
.16
07
0
.11
0
.19
0
.10
0
.12
0
.29
0
.10
1
0
.23
08
0
.09
0
.32
0
.19
0
.11
0
.26
0
.16
0
.24
1
(Sou
rce:
Au
th
o
r’
s
o
wn
pro
cess
in
g
)
Figure
2. Ev
ol
ution Co
nver
ge
nce Time
U
si
ng
4
-
Test
bed
s
Last
ly,
Fig
ur
e
2
sho
ws
e
xec
ution
ti
me
versus
co
nver
ge
nce
usi
ng
f
our
(
4)
s
e
par
at
e
te
st
-
be
ds
to
sim
ulate
the
ef
fecti
ve
ne
ss
an
d
e
ff
ic
ie
nc
y
of
t
he
mode
l.
With
data
lo
gs
of
it
ems
pu
r
chase
d
by
cust
om
e
rs,
t
he
pro
po
s
ed
model
-
based
s
olu
ti
on
co
nver
ges
f
ast
e
r
as
it
ems
co
mm
on
l
y
place
d
in
a
ba
sk
et
are
sel
ec
te
d
[
63
-
64
];
an
d
th
us
,
yields
an
ef
fe
ct
ive
mea
ns
to
maximize
t
he
int
erest
in
gness
of
pro
duct
placeme
nt
on
the
s
helve
s.
T
his
interest
ingness
val
ue(
s
)
a
re
r
ules
mi
ned
by
asso
ci
at
ion
usi
ng
the
fr
e
que
ncy
grow
t
h
pa
th
-
al
gorith
m
f
or
it
em
locat
ion
on
the
she
lves
[
65
-
69].
The
rati
on
al
e
f
or
intere
sti
ng
ness
ma
ximiza
ti
on
with
locat
i
on
c
onsiderati
ons
is
base
d
on
the f
a
ct
that,
associa
ti
on
r
ule
mini
ng
hel
ps
to maxi
mize
cro
ss
-
sel
li
ng
ef
fect
of
it
ems.
Als
o,
it
is
cl
ear
tha
t
the
locat
io
n
of
s
helves
ha
s
the
un
den
ia
bl
e
impact
on
th
e
sel
li
ng
rate
of
it
ems.
T
hus,
it
ems
that
are
pl
aced
near
t
he
entr
ance
or
e
xit
door
s
ha
ve
m
ore
cha
nce
to
be
purc
hased.
S
o,
pr
e
fer
e
nce
functi
on
of
the
s
tore’s
mana
ger
de
pe
nds
on
t
he
fo
ll
owin
g
par
a
mete
rs:
sel
li
ng
benefit
,
s
upport
a
nd
c
onfide
nce
of
eac
h
pair
of
i
te
ms,
and the
sel
li
ng
po
s
sibil
it
y
of e
ach ite
m
from
each s
helf.
4.
CONCL
US
I
O
N
Our
memet
ic
(
gen
et
ic
al
gorit
hm
trai
ne
d
neural
net)
m
od
el
as
us
ed
for
cl
assifi
cat
ion
of
mar
ket
bas
ket
data
–
a
dap
ts
GA
t
o
hel
p
spe
ed
up
the
fin
al
sta
ges
of
A
NN
a
nd
th
us
,
yield
a
r
obus
t
op
ti
ma
in
t
he
sh
ort
es
t
amo
un
t
of ti
me
f
or s
uc
h
a
dyna
mic an
d
c
omp
le
x
ta
s
k. T
he
r
ule
-
based he
uri
sti
cs w
il
l help bet
te
r
re
presen
t data
values
in
the
m
od
el
[
70
-
72].
T
hough,
hybri
ds
are
quit
e
dif
ficult
to
imple
me
nt,
e
xploit
a
nd
exp
l
or
e
–
it
ho
wev
e
r
yields
bette
r
s
olu
ti
ons
with
appr
opriat
e
pa
rameter
sel
ect
ion
that
m
us
t
be
e
nc
od
e
d
t
hro
ugh
the
m
odel
’s
structu
re
d
le
ar
ning.
This
will
in
t
urn
hel
p
a
ddress
the
issu
es
of
sta
ti
sti
cal
de
pe
nd
e
ncies
imposed
on
hy
br
i
d
by
the
under
l
ying
stochastic
he
uri
sti
cs
adopted
,
reso
l
ve
c
onflic
ts
imp
os
e
d
with
e
nc
od
i
ng
of
dataset
c
um
da
ta
feats
of
i
nterest
[
73
]
as
well
as
hi
gh
li
ght
the
i
mp
li
cat
ion
s
of
su
c
h
m
ulti
-
ag
ent
m
od
el
as
agen
ts
seek
t
o
create
their
ow
n
behavio
ur
al
ru
le
s
on
t
he
dataset
us
e
d
–
as
the
model
pr
opose
s
a
so
l
ution
to
dis
play
un
de
rlyin
g
pro
bab
il
it
ie
s
of
data
feats
of
i
nterest.
GA
hel
ps
to
yiel
d
bette
r
ge
nerat
ion
vi
a
it
s
pr
oces
s
of
rec
ombinati
on
an
d
mu
ta
ti
on as a
ppli
ed.
REFERE
NCE
S
[1]
Aguinis,
H,
Fo
rcu
m,
L
.
E,
and
Joo
H.
"
Us
ing
ma
rk
et
b
aske
t
analysis
in
m
ana
ge
me
nt
rese
arc
h,
"
Journal
of
Manage
ment
,
vo
l.
39(7)
,
pp
.
179
9
-
1824,
2013
.
[2]
Prithi
viraj,
P
an
d
Porkodi,
R.
"
A
com
par
ative
ana
lysis
of
asso
ci
a
ti
on
ru
le
mi
n
ing
a
lgori
th
ms
i
n
data
mi
ning
,
"
Ame
rican J. of
C
omputer
Scienc
e
and
Eng
ineerin
g
Surve
y
,
vo
l.
3(
1),
pp
.
98
-
119
,
2
015
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IS
S
N
:
2252
-
8776
In
t J
Inf & C
ommu
n Tec
hn
ol
,
V
ol.
8
,
No.
3,
Dec
201
9
:
12
8
–
138
136
[3]
Barbu,
A
.
"
E
ight
contempor
ary
t
r
ends
in
the
m
ark
et
r
ese
ar
ch
industry,
"
Manag
eme
nt
and
Mar
ke
t
in
g
,
vol
.
8(3
),
pp.
429
-
450,
2013
.
[4]
Ojugo,
A.A
and
Chie
me
k
e,
S
.
"
Inve
ntory
m
ana
ge
me
nt
using
th
e
ma
rke
t
baske
t
an
al
ysis:
a
com
par
at
iv
e
rev
i
ew
of
f
p
-
growth
and
ap
ri
ori
al
gori
thm
,
"
J
ournal
of
Computer
S
ci
en
ce
Ap
pli
cation
–
N
iger
ian
Compute
r
S
oci
e
ty
,
vo
l.
56(4
),
pp
.
21
-
34
,
2019
.
[5]
Coscia
,
C,
Fonta
na,
R
,
and
Sem
er
aro
,
P.
"
Mark
et
Basket
An
al
ysis
for
studying
cu
ltural
Consum
er
B
eha
viour
:
AM
TP
Card
-
Holder
s,
"
Stat
isti
ca Appl
i
c
ata
,
vo
l. 26(2), p
p73
-
92,
2016
.
[6]
Cavi
que
,
L
.
"A
s
ca
l
abl
e
a
lgori
th
m
for the
ma
rk
et ba
sket
analysis,
" Journal
of
R
eta
il
ing and
Consu
me
r
Serv
ic
es
,
vo
l.
14(6),
pp
.
400
-
4
07,
2007
.
[7]
Kaur,
M
and
Ka
ng,
S.
"M
ark
et
Basket
Ana
lysis:
Ide
ntify
th
e
Ch
angi
ng
T
ren
ds
o
f
Marke
t
Data
Us
ing
As
socia
ti
on
Rule
Mining
,
"
Inte
rnational
C
onfe
renc
e
on
Comp
utat
ional
Mode
ll
ing
and
Sec
urit
y
(CM
S
2016)
,
vo
l.
85,
pp78
-
85,
2016
.
[8]
Heyda
ry,
M
and
Yous
efl
i,
A.
"A
new
opt
im
i
za
t
io
n
mod
el
for
ma
r
ket
baske
t
ana
ly
sis
with
allocatio
n
consid
era
t
ions:
A
genetic
al
gori
thm
solu
ti
on
ap
proa
ch,
"
Man
ag
em
en
t
&
Marke
ti
ng.
Cha
ll
eng
es
for
th
e
Know
ledge
Soc
ie
ty
,
vol
.
12(1),
pp
.
1
-
11
,
2017.
[9]
Ojugo,
A
.
A.
"
Uns
uper
vised
cl
us
te
r
pre
d
ic
t
ion
m
odel
for
m
ark
et
baske
t
an
al
ysis:
qu
est
for
bette
r
a
lterna
t
ive
s
t
o
associa
t
ion
rule
mi
ning
appr
o
ac
h
,
"
In
Proc.
o
f
th
e
21st
ISTE
AMS
Conf
.
on
Int
el
l
i
gent
Syst
ems
an
d
Manage
m
ent
o
f
Data
,
pp
.
121
-
1
35,
Re
trieve
d
Se
p
24
2018
.
[10]
Ojugo,
A.A,
Yor
o,
R.
E
,
and
Oko
bah,
I
.
P
.
"
A
compara
t
ive
study
of
ma
rke
t
baske
t
a
nal
ysis
rule
mi
ni
ng
appr
oac
h
es,
"
Techni
ca
l
R
ep
ort
of
th
e
F
ede
ral
Uni
ve
rs
it
y
o
f
P
et
rol
e
um
Re
sour
ce
s
Ef
f
urun
(
FUPR
E
-
TRON
-
102
3)
,
pp
.
312
-
337
,
20
18.
[11]
Dhae
nens,
C
an
d
Jourdan
,
L.
"
Meta
heur
isti
cs
a
nd
As
socia
t
ion
Rule
s,
in
Dhae
nes,
C
.
and
Jou
rda
n,
L
.
(Eds.
)
,
Meta
heur
isti
cs
f
or
Big
Da
ta,
"
Jo
hn
Wiley
&
Sons
,
Hobok
en
,
pp.
8
7
-
108,
2016
.
[12]
Ojugo,
A.A
and
Yoro,
R.
E
.
"
In
t
el
li
g
ent
li
ghtw
eight
ma
rk
et
bask
et
associ
ative
ru
le
m
ini
ng
for
smart
phon
e
c
lou
d
-
base
d
applicatio
n
to
ea
se
b
anking
tra
nsac
ti
ons,
"
In
Proc.
of
the
21st
ISTEAMS
Conf.
on
Int
el
l
i
gent
Syst
ems
an
d
Manage
ment
o
f D
ata
,
pp
.
23
-
37
,
2018.
[13]
Ta
ng,
K,
Chen
,
Y
and
Hu,
H.
"Context
-
b
ase
d
ma
rke
t
baske
t
a
nal
ysis
in
a
mu
l
ti
ple
-
stor
e
envi
r
onme
nt
,
"
Dec
ision
Support
Syste
ms
,
vol
.
45
(1)
,
pp
.
150
-
163,
2008
.
[14]
Chen,
Y,
Ta
ng
,
K,
Shen,
R,
and
Hu,
Y.
"
Marke
t
baske
t
an
al
ysis
i
n
a
multipl
e
stor
e
envi
ron
me
nt
,
"
Dec
ision
Suppor
t
Syste
ms
,
40
,
pp
.
339
-
354,
2005
.
[15]
Yun,
C
,
Chuang
,
K,
and
Chen
,
M.
"
Adhere
nc
e
cl
uster
ing:
an
ef
fic
i
ent
m
et
hod
f
or
mi
ning
ma
rk
e
t
-
baske
t
cl
ust
ers,
"
Information
Syst
ems
,
31,
pp
.
170
-
186,
2006
.
[16]
Russ
el
l,
R.
A
and
Urban,
T
.
L
.
"
Th
e
lo
cation
and
a
llocation
of
produ
ct
s
and
produ
c
t
f
am
ilies
on
r
et
a
il
shelve
s,
"
Annuals
of
Operation
R
e
search
,
vol
.
179
(
1),
pp
.
131
-
147
,
2010.
[17]
Nier
op,
E,
Fok,
D
and
Franses
,
P.
"
Inte
r
action
b
et
wee
n
shelf
l
ay
out
and
m
ark
e
ting
eff
ec
t
ive
n
ess
and
it
s
im
p
ac
t
o
n
opti
mizing
shelf
arr
angeme
nts
,
"
Marke
ti
ng
Scien
ce
,
vol
.
27(6)
,
p
p
.
106
5
-
1082
,
2
006
.
[18]
Sara
f,
R
and
Pat
il
,
S.
"
Marke
t
-
B
aske
t
Analysis
u
sing
Agglom
era
t
ive
Hier
arc
hi
cal
appr
o
ac
h
for
c
l
usteri
ng
a
re
ta
i
l
it
em
s
,
"
Int
ernational
Journal
of
Computer
Scien
ce
and
Ne
twork Sec
urit
y
,
vol.
16
(3),
pp
.
47
-
56
,
2
016.
[19]
Shiokawa,
Y
,
M
isawa
,
T
,
Dat
e
,
Y,
and
Kikuchi,
J.
"
Appli
ca
t
ion
of
ma
rk
et
baske
t
analysis
for
the
visual
i
zation
o
f
tra
nsac
ti
on
d
ata
base
d
on
hum
an
lifestyl
e
and
spec
troscop
ic
me
asure
me
nts
,
"
Analytical
chem
istry
,
vol
.
88(
5),
pp
.
2714
-
2719
,
2016
.
[20]
Solnet
,
D
,
Bozt
u
g,
Y
and Doln
ic
a
r,
S.
"
An
unta
pp
ed gold m
ine? E
xplori
ng
th
e pot
e
nti
al
of m
ark
et
b
aske
t an
al
ysis
to
grow hot
el re
v
en
ue,
"
In
te
rnation
al
Journal
of
Hos
pit
ali
t
y
Manag
e
ment
,
56
,
pp
.
11
9
-
125,
2
016
.
[21]
Holý,
V,
Sokol,
O,
and
Černý
,
M
.
Cluste
ring
Reta
il
Produc
ts
Base
d
on
Customer
Be
havi
or
,
Appl
ie
d
Soft
Comput
ing
,
In
Press
,
Acc
epted
Manuscrip
t,
2
017
.
[22]
Shrivasta
va
,
S
a
nd
Raj
put
,
V
.
"
E
volut
iona
ry
al
go
rit
hm
-
b
ase
d
association rul
e mi
n
i
ng: A
brie
f s
urve
y,
"
Inte
rnationa
l
Journal
of
Innov
ati
on
in
Eng
inee
ring re
search
an
d
Manage
ment
,
vol.
2(1)
,
pp
. 1
-
7,
2015
.
[23]
Meisel
,
S
and
Matt
f
el
d,
D.
"
Syne
rgie
s
of
Oper
at
io
ns
Resea
r
ch
and
Data
Mining,
"
E
uropean
Journal
of
Operationa
l
Re
search
,
vol
.
2
06(1),
pp
. 1
-
10
,
2010.
[24]
Sadh,
A.S
and
Shukla,
N
.
"A
ss
oci
at
ion
rul
es
opt
i
mi
z
at
ion:
A
surv
ey,
"
In
te
rnation
al
Journal
of
Ad
vanc
ed
Compute
r
Re
search
,
vol
.
3
(9),
pp
.
111
-
115
,
2013.
[25]
Sharma
,
A
and
T
iva
ry,
N
.
"
A
surv
ey
of
association
rule
mi
n
ing
usin
g
gen
etic
a
lgori
t
hm,
"
Inte
rnat
ion
al
J.
o
f
Compute
r
Appl
ic
a
ti
o
ns
&
I
nformation
Tech
nology
,
vol
.
1(2)
,
pp
. 5
-
11
,
2012
.
[26]
Ti
wari
,
R
and
Singh,
M.P
.
"
Corr
el
a
ti
on
-
base
d
At
tri
bute
Selecti
on
using
Gene
tic
Algorit
hm
,
"
Inter
nati
onal
Journ
al
of
Computer
Ap
pli
cation
,
vol
.
4(
8),
pp
.
28
-
34
,
20
10.
[27]
Tom
ar
,
N
and
M
anj
hvar
,
A.K.
"
A sur
vey
on
dat
a
mi
ning
op
ti
m
izati
on
te
chn
ique
s,
"
Inte
rnational
J
ournal
of
Scienc
e
Technol
ogy
&
E
ngine
ering
,
vol
.
2(6),
pp
.
130
-
13
3,
2015
.
[28]
Aw
odel
e,
O,
Iz
a
ng,
A.A
,
Kuyoro
,
and
S
Os
isanwo
,
F.
"
Big
Da
ta
an
d
Cloud
Comput
ing
Iss
ues,
"
Inte
r
nati
onal
Journal
of
Computer
Ap
pli
cations
0975
-
8887
,
vol
.
133(1
2),
2016
.
[29]
Annie,
L
and
Kumar
,
A.
"
Mark
et
baske
t
analysi
s
for
a
Superm
a
ke
t
Based
on
Fr
eque
nt
i
tem
Min
ing,
"
Inte
rnat
ion
al
Journal
of
Computer
Sc
ie
nc
e
,
vol
.
9
(5)
,
pp
.
34
-
43
,
2012
[30]
Agrawal
,
R
,
I
mieli
nski
,
R
,
and
Sw
am
i, H.
'
Minin
g
association
b
etw
ee
n
sets
of
ite
ms i
n
m
assive
d
at
ab
ase
s,
"
In Pr
oc
.
of
ACM
SIGMO
D Conf
ere
nc
e
o
n
Manage
ment
o
f
Data
,
pp
.
207
-
216,
Chi
ca
go
,
W
ashingt
on
DC,
U
SA
,
1993
.
[31]
Agrawal
,
R
and
Srikant
,
R.
"
Fast
al
gori
thm
s
for
mi
ning
a
ss
oci
ati
on
rule
s
in
la
rg
e
dat
ab
ase
s,
"
Proc
.
of
20
th
Int
.
Con
f
.
on
Ve
ry
Lar
ge
D
atabases
,
pp
.
48
7
-
499
,
1994
.
Evaluation Warning : The document was created with Spire.PDF for Python.
In
t J
Inf & C
ommu
n Tec
hn
ol
IS
S
N:
22
52
-
8776
In
ve
nto
ry
pred
ic
ti
on
an
d man
ag
e
me
nt in
Ni
ge
ria
us
in
g mar
ket
b
ask
et
an
alysis …
(
Ar
nold
Ad
i
mab
ua O
j
ugo
)
137
[32]
Vija
ya
ta
,
W
and
Deba
jyot
i
,
M
.
"
Mobile
Ag
ent
b
ase
d
Mark
et
Bas
ket
An
al
ysis
on
Cloud,
"
Ind
ia.
2
014
Int
ernati
ona
l
Confe
renc
e
on
I
T.
(pp.
49
-
62)
.
Pune
,
India:
De
partment
Of
Inf
orm
ati
on
Techn
ology
,
Maharas
htra.
Insti
tute
of
Technol
ogy
Pun
e,
Ind
ia
,
2014
.
[33]
Coel
lo
,
C
,
Pulid
o,
G
,
and
L
ec
hug
a,
M
,
"H
andling
mul
ti
p
le
objecti
v
es
with
p
article
s
warm
opt
i
miza
tion,"
Proc
ee
ding
s
of
E
vol
ut
ionary Com
puti
ng
,
vo
l.
8
(3)
,
pp
.
256
-
27
9,
2004
.
[34]
Ojugo,
A,
Ebok
a,
A,
Okonta
,
E
,
Yoro,
R
,
and
A
ghware
,
F.
"
GA
rule
-
base
d
int
rus
ion
detec
ti
on
sys
te
m
,
"
Journal
o
f
Computing
and
I
nformation
Syst
e
ms
,
vol.
3(8), pp
.
1182
-
1194,
201
2
.
[35]
Ojugo,
A.A
and
Yoro,
R
.
"
Co
mput
ational
in
telli
gen
ce
in
sto
c
hasti
c
soluti
on
f
or
Toroi
d
al
Qu
e
en,
"
Progress
in
Inte
lligen
ce Co
mputing
and
Ap
pli
cations
,
vo
l.
2
(1),
pp
.
46
-
56
,
2
013.
[36]
Ojugo,
A.A,
Emudianughe,
J,
Yoro,
R.
E
,
Okont
a,
E.
O
,
and
Ebo
ka,
A
.
"
Hybrid
n
eur
al
net
work
g
r
avi
t
at
ion
al
se
arch
al
gorit
h
m
for
r
ai
nfa
ll
runoff
mode
li
ng
,
"
Pro
gress
in
Intelligence
Comput
i
ng
and
App
li
c
ati
on
,
vol
.
2(1
),
pp.
22
-
33
,
2013
.
[37]
Ojugo,
A
.
A,
Ben
-
Iwhiwhu,
E
,
Ke
kej
e
,
D
.
O,
Yerok
un,
M.O
,
and
Iya
wa,
I
.
J.B
.
"M
a
lwar
e
propa
g
at
ion
on
time
va
rying
net
work,"
In
t. J. Modern E
du
.
Co
mp.
Sci
.
,
vol
.
8
,
pp
.
25
-
33
,
2014
.
[38]
Ojugo,
A.A,
Eb
oka,
A.O,
Yoro
,
R.
E,
Ye
rokun,
M.O,
and
Efo
zi
a
,
F.N.
"
Hybrid
mode
l
for
ea
rly
dia
be
te
s
dia
gnos
is,
"
Mathe
mati
cs
and
Computers
in
Sci
en
ce
and
Ind
ustry
(A
Math
e
matic
s
and
Computers
in
Scienc
e
and
Engi
n
ee
ri
ng
Serie
s)
,
s
eri
es
50
,
pp
.
176
-
182
,
2
015
.
[39]
Fausett
L.
Fund
amental
s of
Neu
ral Ne
tworks
,
New Jersey: Prentice Hal
l
,
pp
.
240
,
1994
.
[40]
Bishop,
C
.
M.
Neural
n
et
wo
rks
for
pa
ttern
rec
ogn
it
ion
,
Oxford
Univer
s
it
y
Press
,
ISB
N:
0
-
342
-
4523
0,
pp.
253
-
294
,
19
95.
[41]
Guo,
W
.
W
and
Xue,
H.
'
An
incorporat
iv
e
sta
ti
st
ic
and
neur
al
ap
proa
ch
for
cro
p
yie
ld
modelli
ng
and
fore
ca
st
ing,
"
Neural
Computi
ng
and
App
li
ca
tions
,
vol.
21(1)
,
pp
.
109
-
117
,
20
11.
[42]
Hall
,
M.
J
.
"
How
well
do
es
you
r
model
f
it
d
ata?
,
"
Journal
of
H
y
droinformatic
s
,
vol.
3(1)
,
pp
.
49
-
55,
2001
.
[43]
Haykin,
S
.
Neur
al
ne
tworks,
comprehe
nsive f
oun
dati
on
,
Pren
tice
Hall
:
New Jerse
y,
pp
.
23
-
8
9,
19
99
.
[44]
Heppne
r,
H
and
Grena
nd
er,
U.
Stoc
hastic
non
-
l
ine
ar
model
for
coordinat
ed
bir
d
fl
o
ck
s,
In
Kra
sner,
S
(
Ed.
)
,
T
he
ubiqui
ty of
chao
s
,
pp.
233
-
238,
Wa
shington:
AA
AS
,
1990.
[45]
A.A.
Ojugo
and
A.O.
Ebok
a.
"
Si
gnat
ure
-
b
ase
d
m
al
ware
de
tecti
on
using
appr
oximate
Boyer
Moor
e
string
m
atchin
g
al
gorit
h
m,
"
Int.
J.
o
f
Ma
the
mati
c
al
Sc
ie
nc
es
and
Computing
,
vo
l.
3(5),
pp
.
49
-
62
,
2019.
[46]
R.
E.
Yoro
and
A.A.
Ojugo.
"
A
n
intelligent
mo
del
to
pr
edi
c
t
th
e
re
lationship
o
f
wea
th
er
conditi
ons
for
fish
far
m
ing
produc
ti
on
yield
in
Nig
eria,
"
Ame
rican Journal
o
f
Mode
li
ng
and
Optimizati
on
,
vo
l.
4(3)
,
pp
.
79
-
90
,
2019
.
[47]
R.
E.
Yoro
and
A.A.
Ojugo.
"
Que
st
for
pre
valenc
e
rat
e
of
Hepa
t
it
is
-
B
Virus
infe
ctio
n
in
Niger
i
a:
co
mpa
ra
ti
ve
study
of
supervise
d
ve
rsus
unsupervise
d
model,
"
Ame
ri
can
J
.
o
f
Mod
el
i
ng
and
Optimiza
ti
on
,
vol. 4(3)
,
pp
.
79
-
90
,
2019
.
[48]
A
.
A.
Ojugo
and
A.O.
Ebok
a.
"
Modeli
ng
th
e
c
omput
ationa
l
soluti
on
of
ma
rk
e
t
baske
t
associative
ru
le
mi
nin
g
appr
oac
h
es
usin
g
dee
p
neur
al network,
"
Digital
Technol
ogi
es
,
v
ol.
3(1)
,
pp
. 1
-
8,
2018.
[49]
A.A.
Ojugo
and
A.O.
Ebok
a.
"
Co
mpa
ra
ti
ve
ev
al
u
a
ti
on
for
per
formance
ada
p
ti
ve
mo
del
for
spam
phis
hing
de
tecti
on
,
"
Digit
al
Te
chnol
o
gie
s
,
vo
l. 3(1),
pp
.
9
-
15
,
2018
.
[50]
Pere
z,
M
and
M
arwa
la,
T.
"
Stoc
hasti
c
op
ti
m
ization
appr
oa
che
s
f
or
solving
Sudo
ku,
"
IEEE
Tr
ansacti
on
on
Ev
o
l.
Comp
.
,
pp
.
256
-
279,
2011
.
[51]
Mandic
,
D
and
C
ham
ber
s,
J.
Recurr
ent
N
eural
Net
works
for
Pr
edi
c
ti
on:
Learni
ng
A
lgorit
hms
,
Archi
t
ec
tures
an
d
Stabi
lity
,
W
iley
&
Sons
:
New Y
ork,
pp
.
56
-
90
,
2
001
.
[52]
Ina
n,
G
and
El
if
,
D.
Adapti
ve
n
eur
o
-
fuz
zy
inf
ere
n
c
e
sys
te
m
for
c
la
s
sific
a
ti
on
of
E
E
G
using
wave
le
t
coe
ffi
ci
en
t
2005)
.
R
et
ri
eve
d
f
rom
:
htt
p://ww
w.rorylewis.c
om
/PDF
/04
[53]
Ludm
ila
I
.
K
.
"
Fuzzy class
ifi
ers
,"
Schol
arped
ia
,
3
(1),
pp
2925
,
20
08.
[54]
Urs
em
,
R
,
Krink
,
T,
Jensen,
M,
a
nd
Michalewi
cz,
Z
.
"
Analysis
an
d
modeling
of
co
ntrol
s
in
dyna
mic
sys
te
ms
,
"
IE
E
E
Tr
ansacti
on
on
Me
metic
S
yste
m
s and
Ev
o
lut
iona
ry
Computing
,
v
ol.
6(4)
,
pp
.
378
-
389
,
2002
.
[55]
Minns,
A.
Artif
ic
ial
neural
n
etwor
ks
as
sub
-
s
ymboli
c
proc
ess
desc
riptors
,
p
ubli
shed
PhD
The
sis,
B
al
ke
m
a,
Rott
erd
am
,
Neth
erl
ands
,
1998.
[56]
Reynol
ds,
R.
"
In
troduc
ti
on
to
cultura
l
a
lgori
th
ms,
"
Tr
ansacti
on
on
E
vol
ut
ionary
Pr
ogram
ming
(IE
E
E)
,
pp.
131
-
139
,
1994.
[57]
Olson,
D.L
.
"
Marke
t
Baske
t
An
al
ysis,
Chapter
Descri
pti
v
e
Da
t
a
Mining
,
"
Part
of
th
e
seri
es:
C
omputati
onal
Ris
k
Manage
ment
,
pp
.
29
-
4
1.
2016
.
[58]
Wa
iswa,
P.P.W
and
Bary
am
ur
e
eba
,
V.
,
"
Ext
r
action
of
int
er
esting
associa
t
ion
r
ule
s
using
Gene
ti
c
Algor
it
hms
,
"
Inte
rnational
Jo
urnal
of
Comput
ing
and
ICT
Re
s
earc
h
,
vo
l.
2(1),
pp
.
26
-
33
,
2008
.
[59]
Woo,
J
.
"
Mar
k
et
B
aske
t
Anal
ysis
using
Spa
rk,
"
A
RP
N
Jou
rnal
of
S
cienc
e
and
T
ec
hnol
ogy
,
vo
l.
vol
.
4
,
pp.
207
-
209
,
20
15
.
[60]
Mara
fi,
S.
Mark
et
B
aske
t
Analys
is
with
R
,
2014
.
Ret
ri
eve
d
from:
htt
p://ww
w.sal
e
mm
ar
afi
.
com/co
de/
m
ark
et
-
b
aske
t
-
ana
lysis
-
with
-
r
/.
[61]
Rosaria
,
S.
(2
015).
How
to
buil
d
a
Mar
ket
B
aske
t
An
al
ysis
Eng
ine.
Ret
r
ie
v
ed
fro
m
Analy
ticbrid
ge:
htt
p://ww
w.a
nalytic
bridg
e.
d
atasc
ie
nc
ec
en
tral.
com/profiles/blogs/h
ow
-
to
-
buil
d
-
a
-
m
ark
et
-
b
aske
t
-
analysis
-
engi
ne
.
[62]
Mahe
shw
ari
,
N
,
Pandey
,
N.
K
,
and
Agarwa
l,
P.
"
Marke
t
B
aske
t
Analysis
us
ing
As
sociation
Rule
Learni
ng,
"
Inte
rn
ati
onal
Jo
urnal
of
Comput
er
Applications
,
0975
-
8887,
201
6.
[63]
Komal
Khur
ana
and
Mrs
.
Simp
le
Shar
ma.
"
A
Compa
ra
ti
ve
An
al
ysis
of
As
socia
ti
on
Rule
s
Min
ing
Algori
thm
s,
"
Inte
rnational
Jo
urnal
of
S
ci
en
ti
f
i
c
and
Re
search
Publ
ic
a
ti
ons
,
vo
l.
3(
5
)
,
May
201
3
.
[64]
Agrawal
,
R
,
I
mieli
nski
,
R,
and
S
wami
,
S.
"
A
Mi
ning
associ
at
ion
Bet
wee
n
Set
s
of
Ite
ms
in
Mass
iv
e
Dat
aba
ses,"
In
Proc.
o
f ACM
SI
GMO
D Conf
ere
nce
on
Manage
ment
of
Data
,
pp
.
207
-
216
,
1993
.
Ret
r
ie
ved
Sept
e
mbe
r
24
,
2016
Evaluation Warning : The document was created with Spire.PDF for Python.