TELKOM
NIKA
, Vol.11, No
.11, Novemb
er 201
3, pp. 6434
~6
440
e-ISSN: 2087
-278X
6434
Re
cei
v
ed Ap
ril 20, 2013; Revi
sed
Jun
e
24, 2013; Accepted July 1
1
,
2013
A Novel Algorithm of Network Trade Customer
Classification based on
Fourier Basis Functions
Li Xin
w
u
*
1
, Guan Peng
c
h
eng
2
Schoo
l of Internatio
nal T
r
ade
and Eco
nom
ics, Jiang
xi
U
n
iv
ersit
y
of F
i
n
anc
e and Eco
nom
i
cs, Nancha
ng,
Jian
g
x
i, Chi
na,
3300
13
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: li
yue
751
1@
1
63.com
1
, 407
7
064
83@
qq.co
m
1
A
b
st
r
a
ct
Lear
nin
g
a
l
gor
i
t
hm
of ne
ural
netw
o
rk is alw
a
ys an
i
m
p
o
rta
n
t researc
h
co
ntents in
ne
ura
l
netw
o
rk
theory r
e
searc
h
a
nd
ap
plic
ati
on fi
eld,
lear
ni
ng
alg
o
rith
m
a
bout th
e fe
ed-f
o
rw
ard n
eura
l
netw
o
rk has
n
o
satisfactory sol
u
tion i
n
p
a
rticu
l
ar for its defe
c
ts
in calcu
l
ati
on sp
eed. T
h
e
pap
er pres
ent
s a new
F
ouri
e
r
basis
functi
on
s ne
ural
n
e
tw
ork a
l
gor
ith
m
and
a
ppl
ied
it
to cl
assify
n
e
tw
ork trade
c
u
stomer. F
i
rst, 21
customer clas
sificatio
n
ind
i
c
a
tors are d
e
si
gne
d, base
d
on char
acteris
t
ics and b
eha
viors an
alysis
of
netw
o
rk trad
e
customer, i
n
cl
udi
ng
custo
m
e
r
char
acte
ristic
s type v
a
ri
abl
e
s
an
d c
u
sto
m
e
r
be
havi
o
rs ty
p
e
varia
b
les. Sec
ond, F
our
ier
b
a
sis functi
ons i
s
used to
i
m
pr
ove the c
a
lc
ula
t
ion flow
an
d a
l
gorit
hm structu
r
e
of orig
ina
l
BP
neur
al n
e
tw
ork alg
o
rith
m to s
pee
d u
p
its co
nverg
ence
an
d
then a
new
F
ouri
e
r bas
is ne
ura
l
netw
o
rk mod
e
l
is construct
e
d. F
i
nal
ly the
exper
imenta
l
r
e
sults sh
ow
that the pr
obl
e
m
of conv
erge
n
c
e
spee
d can
be
en solv
ed, a
n
d
the accur
a
c
y
of the
custo
m
er cl
assific
a
ti
on are
ensur
e
d
w
hen the n
e
w
alg
o
rith
m is us
ed in n
e
tw
ork trade custo
m
er classificati
on p
r
actically.
Ke
y
w
ords
:
BP
neura
l
netw
o
r
k
algor
ith
m
, F
ourier b
a
sis fun
c
tions, custo
m
er classific
a
tio
n
, calcul
atio
n fl
ow
,
alg
o
rith
m struc
t
ure
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Artificial n
eural net
work i
s
a h
o
t re
se
a
r
eh fi
el
d in
rece
nt years,
involving el
ectro
n
ic
sci
ene
e and
techn
o
logy, informatio
n a
nd co
mmuni
cation engi
ne
ering,
comp
u
t
er scien
c
e a
nd
techn
o
logy, control scie
nce and techno
logy, and ma
ny other disci
p
line
s
. The a
pplication fiel
ds
of artificial n
eural
network in
clud
e m
odelin
g,
time seri
es
anal
ysis, pattern
reeo
gnition
and
control, and
many other fi
elds related.
And for E
-
co
mmerce
ente
r
pri
s
e
s
, a
s
th
ere
are
vast
cu
stome
r
s i
n
netwo
rk tra
n
s
a
c
tion,
these
cu
stom
ers
differ in thou
san
d
s of
ways. Fo
r dif
f
erent custo
m
ers,
their d
e
mand
s a
r
e
ever
cha
ngin
g
. It’s impo
ssi
ble fo
r E-comm
erce ente
r
pri
s
e
s
to meet the
d
e
mand
s
of all
the custo
m
ers,
whi
c
h i
s
not
only limited b
y
self mate
ri
al co
nditi
on
s
of enterpri
s
e
s
, but al
so
u
nde
sira
ble in
the
asp
e
ct of e
c
onomi
c
be
ne
fits. Therefo
r
e, E-co
m
m
erce ente
r
p
r
ise
s
shall pi
ck out the mo
st
valuable cu
st
omers whom
they
can effectively
se
rve; instea
d of
hitting out i
n
all directio
ns,
enterp
r
i
s
e
s
shall p
r
ovide t
hem
with mo
re in
divi
dual
servi
c
e, a
nd
give co
nsi
deration to e
a
ch
transactio
n
custome
r
. So
to corre
c
tly and effect
ivel
y classify tra
n
sa
ct
ion cu
stomers
play
s a
signifi
cant
rol
e
for ente
r
p
r
ise
s
to
carry out in
dividu
al servi
c
e
a
nd m
a
rketing
strategie
s
f
o
r
different cu
st
omers [1].
2. Literature
Rev
i
e
w
The
widely
-
u
s
ed
metho
d
s of ente
r
p
r
ises fo
r cu
sto
m
er cla
s
sification
at pre
s
ent
are
mainly qu
alitative method
and
qua
ntitative method
. As the
qua
litative metho
d
for
cu
stom
er
cla
ssif
i
cat
i
on
is ju
st
t
o
cla
s
sif
y
all t
h
e
t
a
r
get
c
u
st
o
m
er
s of
ent
e
r
pri
s
es i
n
t
h
e
ma
cro
s
copi
c l
e
v
e
l,
cu
stome
r
cla
ssifi
cation i
s
carried o
u
t according t
o
different value emp
h
a
s
is of different
cu
stome
r
s. T
he form
ation
of cu
stome
r
value
is
si
mply expre
ssed a
s
: Value
= Benefit-Cost.
Qualitative cl
assificatio
n
m
e
thod
cla
ssifi
es
cu
stome
r
s in a si
mple
way, only offeri
ng gui
dan
ce f
o
r
cu
stome
r
cla
ssifi
cation
of enterp
r
i
s
e
i
n
the
ma
croscopic level, u
nable
to p
r
ov
ide
spe
c
ific a
nd
c
r
ed
ib
le b
a
s
i
s
fo
r
e
n
t
er
pr
is
e d
e
c
i
s
i
o
n
s
;
fu
r
t
h
e
r
m
or
e
,
as
th
er
e is
no
s
t
r
i
ct p
r
oc
es
s
of
argu
mentatio
n, the m
e
tho
d
de
pen
ds o
n
de
cid
e
r’
s
subje
c
tive infe
ren
c
e, th
ere
may be
cert
ain
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Novel Alg
o
rithm
of Network T
r
ad
e Cu
st
om
er Cla
ssifi
cation b
a
sed
on Fou
r
ier… (Li Xinwu)
6435
deviation
s in the analysi
s
pro
c
e
ss, e
a
si
ly resulti
ng i
n
faulty decisi
ons. Fo
r this
rea
s
on, to tru
l
y
provide
cust
omer cla
s
sification
information b
ene
ficial to e
n
terp
rises
sh
o
u
ld de
pen
d
on
quantitative tech
nolo
g
y for custo
m
er
cla
ssifi
cation [2, 3].
Quantitative
cla
ssifi
cation
method i
s
to
apply qu
antitative analysis te
chn
o
logy to
con
d
u
c
t cu
stomer
cla
ssifi
cation on the basi
s
of som
e
spe
c
ific
customer varia
b
l
e
s (credit level of
cu
stome
r
s, p
u
rcha
sing
po
wer of
custo
m
ers, characteri
stics
of d
e
mand
of
c
u
st
ome
r
s,
et
c.
).
Curre
n
tly, there are mainly
two cate
go
rie
s
of dat
a mini
ng for qu
antitative custo
m
er cl
assificati
on
resea
r
ch, whi
c
h
are
tra
d
itional
statisti
ca
l method
an
d
non
-stati
stical metho
d
. T
he former ma
inly
inclu
d
e
s
clu
s
ter analy
s
is,
Bayesian
cla
ssifi
cation,
fa
ctor a
nalysi
s
method, et
c.; this statisti
cs-
based m
e
tho
d
is un
able
to process
a
great
deal
of
sop
h
isti
cated
cu
stom
er
da
ta, and th
ere
are
some
p
r
obl
e
m
s on
th
e accuracy
of cu
stomer cla
s
sifi
cation
results, so
to fun
d
a
m
entally
solv
e the
probl
em of
cu
stome
r
cl
a
ssifi
cation
ne
eds to
re
ly
on no
n-statistical cu
stom
er cl
assification
method, which mainly in
cl
ude
s neu
ral
netwo
rk, fu
zzy set method
, associ
ation
rule
s, gen
etic
algorith
m
, etc. The cla
ssifi
cation te
chn
o
l
ogy bas
ed o
n
neu
ral net
work i
s
com
b
ined with
ce
rtain
informatio
n techn
o
logy, whi
c
h is
a kin
d
o
f
mat
hematical method a
p
p
lica
b
le to co
mplex variabl
es
and multi influen
cing fa
cto
r
s
calculation
,
so it is
more effective in solving com
p
lex cu
stome
r
cla
ssifi
cation
probl
em
s wit
h
better cla
ssification
a
c
curacy, however, the converg
ence pro
b
lem
of
the function
itself greatly limits its applicatio
n valu
e in spe
c
ific proje
c
t p
r
a
c
tice. Seco
ndl
y,
cla
ssifi
cation
is mainly b
a
sed on
su
ch
mathemat
i
c
al
method
s a
s
fuzzy
clu
s
te
ring, roug
h set,
asso
ciation rules, etc., althoug
h these
method
s offer cl
assificati
on rea
s
o
n
e
x
planation in
a
relat
i
v
e
ly
cl
ea
r w
a
y
wit
h
be
t
t
e
r cla
s
sif
i
cat
i
on r
e
s
u
lt
s
un
der t
h
e
cir
c
u
m
st
an
ce
s of
sat
i
sf
a
c
t
o
ry
d
a
t
a
con
d
ition
s
, the modelin
g proce
s
s nee
ds
to provide
sp
ecific m
a
the
m
atical e
quat
ions. As a
re
sult,
these meth
od
s are limite
d
by data con
d
i
t
ions in
spe
c
ific appli
c
ation,
always h
a
ving probl
em
s like
insuffici
ent cl
assificatio
n
a
c
cura
cy or
p
oor “r
o
b
u
s
tn
ess”, limiting
the appli
c
ati
on in custo
m
er
cla
ssifi
cation.
Due to lots of influencin
g
factors
relat
ed to cu
stom
er cla
s
sificati
on, more oft
e
n
than not, the compli
cate
d relation
s are d
i
fficult
to be expresse
d in mathemati
c
al
equation
s
[4].
Cu
stome
r
cl
a
ssifi
cation m
o
dels
ba
sed
o
n
data
mini
ng
have hig
h
cl
assificatio
n
a
c
cura
cy
but leaves be
hind the que
stion of
slow converg
e
n
c
e speed of its al
gorithm. The
r
efore, it is hard
to put into effect in cu
stom
er cla
s
sificati
on.
Based o
n
BP neural ne
twork, Fou
r
ier basi
s
functio
n
s
neural n
e
two
r
k is bei
ng
constructe
d
wi
th Fou
r
ie
r b
a
s
is fun
c
tion
s
in this p
ape
r. In
so
doin
g
,
not
only the pro
b
l
em of conve
r
gen
ce
spe
e
d
has b
een
so
lved, but also the simpli
ci
ty of the model
stru
cture and
the accuracy
of t
he classifi
cation a
r
e en
sured.
3. Selection of Cu
stome
r
Classifica
ti
on Indicator
s
The
sel
e
ction
of rea
s
on
abl
e cl
assificatio
n
vari
able
s
i
s
the
basi
s
of
corre
c
t a
nd
e
ffective
cu
stome
r
cla
ssifi
cation, n
a
mely esta
blishin
g
sc
ientif
ic
and rea
s
o
nable cla
ssifi
cation
in
dicators
system. In view of the na
ture of tradin
g
and ow
n chara
c
te
risti
c
s of online trading, this Pa
per
adopt
s cu
sto
m
er cha
r
a
c
te
ristics type v
a
riabl
e and
cu
stome
r
be
haviors type
variable in
the
spe
c
ific
sele
ction of custo
m
er cl
assifica
tion variable
s
[5].
3.1. Selectio
n of Cus
t
om
er Char
ac
ter
i
stics T
y
pe Variable
Cu
stome
r
ch
ara
c
teri
stics type variable
is
mainly u
s
ed fo
r getting the information o
f
cu
stome
r
s’ b
a
si
c attribute
s
. Such varia
b
le indi
cators as geog
ra
phi
cal po
sition, age, sex, income
of individual
cu
stome
r
pla
y
a key
role
in det
e
r
minin
g
the me
mbe
r
s
of so
me
market segm
ent.
This
kind of
variable
s
mai
n
ly come
s from cu
st
ome
r
s’ re
gistration
information
and custo
m
e
r
s’
basi
c
info
rma
t
ion colle
cted
from the ma
n
ageme
n
t sy
st
em of ban
ks, the co
ntents
of whi
c
h mo
st
ly
indicate the
static data
of custome
r
s’ ba
sic attri
bute
s
,
the adva
n
tag
e
of whic
h is
that most
of the
conte
n
ts
of variabl
es a
r
e
easy to
colle
ct. But so
me
of the b
a
si
c
custome
r
-de
s
cribed
content
s of
variable
s
a
r
e
lack of differe
nce
s
at times [2, 3]
Based
on
an
alyzing
and
summari
zin
g
e
x
isting
literature
s
, the
cu
st
om
er charact
e
risti
cs
type variabl
e
s
de
sig
ned
i
n
this
pap
er
inclu
de:
Cust
omer No.,
P
o
st Code, Date
of
Birth, Sex,
Educatio
nal Backgroun
d,
Occup
a
tion, Monthly
In
c
o
me
, T
i
me
o
f
F
i
r
s
t
W
e
bs
ite Br
o
w
s
i
ng
, an
d
Marital Status.
3.2. Selectio
n of Cus
t
om
er Beh
a
v
i
ors
T
y
pe Variables
Cu
stome
r
be
haviors type variable
s
mainly i
ndicate
a serie
s
of variable ind
i
cato
rs
related
to
cu
stome
r
tran
sacting
be
havi
o
r
and
rel
a
tio
n
with
ba
nks,
whi
c
h
a
r
e
used to
define
t
h
e
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 643
4 – 6440
6436
orientatio
n which e
n
terp
ri
ses shoul
d stri
ve for
in som
e
market se
g
m
ent, and are the key fact
ors
for a
s
certaini
ng target m
a
rket. Cu
sto
m
er b
ehav
io
rs type
varia
b
les i
n
cl
ude
the record
s of
cu
stome
r
s
bu
ying
services or pro
d
u
c
ts, reco
rd
s
of
customer
se
rvice
or produ
ctio
n con
s
umptio
n,
conta
c
t re
co
rds
b
e
twee
n
cu
stom
ers and enterpr
i
s
e
s
,
a
s
well
as custo
m
ers’
con
s
umi
ng
behavio
rs, p
r
eferen
ce
s, life style,
and o
t
her rel
e
vant informatio
n [4].
Based o
n
an
alyzing an
d summari
zin
g
e
x
isting literat
ure
s
, the cu
stomer beh
aviors type
variable
s
d
e
signed in thi
s
pape
r in
clud
e
Monthly
Fre
quen
cy of Website
Login,
Monthly We
b
s
ite
Staying Time, Monthly Times of Purcha
sing, M
ont
hly Amount of Purcha
sing, Ty
pe of Con
s
u
m
er
Produ
cts Pu
rcha
se
d, Time
s of Servi
c
e
Feedb
ack,
S
e
rvice S
a
tisfa
c
tion, Cu
sto
m
er Profitabil
i
ty,
Cu
stome
r
Profit, Repe
at P
u
rcha
se
s, Re
comm
end
ed Numb
er
of Custome
r
s,
Pu
rch
a
si
ng Gro
w
th
Rate.
4. Resear
ch
Method
4.1. Working
Principle of BP Neur
al Net
w
o
r
k Algor
ithm
BP neural
netwo
rk alg
o
rithm h
a
s
uique
adv
an
tages to
de
script the
n
on-lin
ea
r
relation
shi
p
and strong f
unctio
n
simul
a
ting ca
pabili
ty.
It not only has input
and outp
u
t-la
yer
node, but al
so hidde
n-l
a
ye
r nod
e. Its hi
dden
-laye
r
n
euro
n
s
ado
pt S type variation fun
c
tion a
n
d
output-laye
r
neuron
s use pure lin
ear
conversion fun
c
tion so that
BP neural ne
twork algo
rith
m
can
be
close
to the co
rre
s
po
ndin
g
rel
a
tionship
bet
wee
n
any fu
nction
s a
nd
data if there
are
enou
gh hid
d
en layers an
d neu
ron
s
th
eoreti
c
ally. T
herefo
r
e, in t
he study of color ma
nag
e
m
ent,
the
ma
pping
relation
s
amo
ng
the
colo
r spa
c
e
s
of
diff
erent
eq
uipm
ents ca
n be derived
throu
gh
the trainin
g
o
f
standa
rd o
u
t
put data and
measur
e
m
e
n
t data to co
mplete their
conve
r
si
on.
BP
neural net
work algo
rithm i
s
gene
rally co
nsi
s
ted of
three layers of n
euro
n
s
as
sh
own in Fi
gu
re
1
[6, 7].
Figure 1. Wo
rking Pri
n
ci
ple
of BP Neural
Netwo
r
k Alg
o
rithm
4.2. Continu
ous-time Fo
urier Series of Periodic Signal
As we all
kn
ow, for
sig
nal
)
(
t
f
that the pe
ri
od is
T
, it can
be sho
w
ed
b
y
continu
o
u
s
-
time Fourier series
, i.e. formula 1 [8].
)
sin(
)
cos(
)
(
0
1
0
1
0
t
n
b
t
n
a
a
t
f
n
n
n
n
(1)
Of formul
a 1,
T
2
0
is funda
me
ntal an
gula
r
f
r
equ
en
cy,
0
a
is DC
comp
on
ent, and
n
n
b
a
、
are Fou
r
ie
r serie
s
, i.e. formula 2.
T
dt
t
f
T
a
0
0
)
(
1
T
n
dt
t
n
t
f
T
a
0
0
)
cos(
)
(
2
T
n
dt
t
n
t
f
T
b
0
0
)
sin(
)
(
2
(2)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Novel Alg
o
rithm
of Network T
r
ad
e Cu
st
om
er Cla
ssifi
cation b
a
sed
on Fou
r
ier… (Li Xinwu)
6437
For time
-limited non
pe
riodi
c si
gnal
)
(
t
f
,
T
t
0
,
the peri
odi
c sig
nal that
)
(
t
f
is vi
a
contin
uation of
period
T
is
)
(
t
f
p
, i.e. formula 3.
m
)
(
)
(
mT
t
f
t
f
p
(3)
Of formula 3,
m
is a po
sitive numb
e
r.
)
(
)
(
t
f
t
f
p
occurs o
b
viou
sl
y when time
t
is
T
t
0
. Therefore, the
contin
uou
s-ti
me
serie
s
of pe
riodi
c
si
gnal
)
(
t
f
p
can
be
al
so
sh
owed
b
y
Formul
a 1 wit
h
in the prin
ci
pal value pe
ri
od
T
t
0
.
For ba
ndlimit
ed sig
nal
)
0
)(
(
N
t
f
,
formula 1 ca
n b
e
cha
nge
d as formula 4.
)
sin(
)
cos(
)
(
0
1
0
1
0
t
n
b
t
n
a
a
t
f
N
n
n
N
n
n
(4)
For the nu
me
rical
com
puta
t
ion, formul
a
4 is se
parate
d
into formula
5.
)
sin(
)
cos(
)
(
0
1
0
1
0
s
N
n
n
s
N
n
n
tkT
n
b
kT
n
a
a
k
f
(5)
Of formula 6,
s
T
is a samplin
g perio
d, and
N
T
N
T
s
2
0
. When
N
T
T
s
2
, formula 5
can b
e
ch
ang
ed as fo
rmula
6.
)
sin(
)
cos(
)
(
1
1
0
nk
N
b
nk
N
a
a
k
f
N
n
n
N
n
n
(6)
In formula 6,
1
2
...
2
,
1
,
0
N
k
.
4.3. Impro
v
ing BP Neu
r
a
l
Net
w
o
r
k
w
i
th Fourier
Basis Func
tio
n
In formul
a 6,
neu
ral
net
work mo
del
b
a
se
d o
n
F
o
u
r
ier ba
si
s fun
c
tion i
s
p
ro
duced if
)
(
k
f
is a neu
ral n
e
twork outp
u
t
,
)
(
t
f
d
is a ne
ural netwo
rk traini
ng sa
mple,
n
n
b
a
、
are neu
ral
netwo
rk trai
ning
wei
ghts, and
)
cos(
nk
N
and
)
(
sin
nk
N
are
ne
ural
netwo
rk ex
ci
tation
function
s. See Figure 2 [9, 10].
The algo
rithm
of neural net
work mo
del b
a
se
d on fouri
e
r ba
sis fou
n
ction is a
s
foll
ows:
1.
See formula
6 for neu
ral n
e
twork outp
u
t.
2.
See formula
7 for error fun
c
tion of network m
odel.
)
(
)
(
)
(
k
f
k
f
k
e
d
(7)
3.
See formula
8 for netwo
rk
model pe
rformance index.
4.
Weig
ht adju
s
tment by gradi
ent desce
nt
algorithm, See
formula 8 an
d 9 for weig
ht
adju
s
tment q
uantity.
N
n
nk
N
k
e
a
J
a
k
n
k
n
...
2
,
1
,
0
,
)
cos(
)
(
(8)
N
n
nk
N
k
e
b
J
b
k
n
k
n
...
2
,
1
,
0
,
)
sin(
)
(
(9)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 643
4 – 6440
6438
5.
See formula
10 and 1
1
for weight adj
ust
m
ent, in which,
is a learni
n
g
rate, and
1
0
.
N
n
nk
N
k
e
a
a
k
n
k
n
...
2
,
1
,
0
,
)
cos(
)
(
1
(10)
N
n
nk
N
k
e
b
b
k
n
k
n
...
2
,
1
,
0
,
)
sin(
)
(
1
(11)
Figure 2. Wo
rkin
g Prin
cipl
e of Fourie
r Basi
s Ne
ural
Network Mo
del
4.4. Solution of Impro
v
ed Algorithm
(1) Netwo
r
k trainin
g
: Th
e net
work t
r
aining
empl
o
y
s BP ne
ural net
work
a
l
gorithm
algorith
m
by
assigni
ng the
value
s
of all
the
cla
ssification indi
cato
rs f
r
om th
e training
datab
a
s
e
as inp
u
t value and that of classification
weight a
s
o
u
tput. In this algorith
m
, bo
th weight val
u
e
and thresh
old
value are ra
ndomly picke
d
out in
the range of -0.5~0.5, with ade
quate adj
ust
m
ent
with reg
a
rd to the real co
nverge
nce.
(2) Initiali
zati
on: to initialize the weig
ht coeffici
ent wit
h
a small ran
dom num
ber.
(3)
Circulatio
n: to set an iteration nu
m
ber
an
d load
data to und
ergo n
e
two
r
k training.
The weight
coeffici
ent re
quire
d is
acquire
d on
ce
the accu
ra
cy of
desig
nat
ed custom
er is
rea
c
he
d.
(4) Ke
ep the
value of weig
ht coeffici
ent
of Fouri
e
r b
a
s
is
neu
ral n
e
t
work an
d co
nclu
de
the training.
4.5. Conv
ergence Analy
s
is of the Impr
ov
ed Model
As we all
know, the si
ze of learni
n
g
rate
affects neu
ral n
e
twork
conve
r
gen
ce
signifi
cantly. If too small, th
e convergen
ce spee
d
of n
eural
net
wo
rk
is slo
w
and
the comp
utation
amount a
n
d
time are i
n
crea
sed; if
too lar
ge,
neural net
wo
rk
sho
c
ks
n
o
t to rea
c
h
th
e
conve
r
ge
nce. For ab
solut
e
convergen
ce
of
neu
ral
network, a
theorem of
neu
ral
network
conve
r
ge
nce is given a
s
be
low.
Only whe
n
th
e learning rate
sat
i
sf
ie
s
1
3
4
0
N
,
neural netwo
rk al
gorith
m
is
conve
r
ge
nt. Here
N
2
is the
number of
neural netwo
rk trai
ning sample
s. For the space
limitation, see
Refere
nce 8 for t
he detaile
d proof of Th
eore
m
1.
5. Results a
nd Analy
s
is
In orde
r to test the effectiveness of improv
e
d
alg
o
rithm in thi
s
thesi
s
, si
mulation
hard
w
a
r
e
is
Dell Po
we
re
d
ge
R71
0
, in
whi
c
h
pro
c
e
s
sor i
s
E55
06,
memo
ry 2
G
, ha
rd
disk
16
0G;
softwa
r
e
plat
form i
s
Win
dows XP
o
peratin
g
syst
em, Matlab
7
10 p
r
og
ram
m
ing la
ngu
a
g
e
environ
ment.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Novel Alg
o
rithm
of Network T
r
ad
e Cu
st
om
er Cla
ssifi
cation b
a
sed
on Fou
r
ier… (Li Xinwu)
6439
5.1. Process
of Experime
ntal Verifica
tion
The process
of the experi
m
ental ve
rification ca
n be l
i
sted a
s
follo
ws.
(1)
wh
at is to
be p
r
o
c
e
s
sed
duri
ng the
cl
assifi
catio
n
is the num
eri
c
data, so
the
n
u
meri
c
codi
ng on
ch
ara
c
ter d
a
ta should b
e
co
n
ducte
d first;
(2) if the val
u
e num
be
r of
certai
n attri
b
ute is
eq
ual t
o
sample
nu
mber, it m
e
a
n
s th
at it
has little effect o
n
cla
s
sification, h
e
n
c
e,
remove
su
ch
attribut
e first. T
h
re
e attribute
s
as
Cu
stome
r
No
., Post Code
and Date of Birth are
remo
ved in this ca
se.
(3) e
s
tablish
trainin
g
sa
mple
set a
c
cording
to d
o
main
(p
rior) kn
owle
dge.
Time
s of
purcha
s
in
g a
nd total amo
u
n
t of purch
asi
ng of ea
ch
cu
st
ome
r
a
r
e t
w
o majo
r f
a
ct
o
r
s
of
cu
st
ome
r
cla
ssifi
cation
(this is the p
r
i
o
r kn
owl
edg
e
of dom
ain), so sele
ct 400
pieces of typical data am
o
n
g
all the
cu
sto
m
ers to fo
rm traini
ng
sample
set.
And divide
them into
fo
ur type
s a
s
Gold
Cu
st
ome
r
s,
S
ilv
er
Cu
st
o
m
er
s,
Ordina
ry Cu
stom
ers, Potentia
l
Cu
stome
r
s a
c
cordi
ng to
ABC
manag
eme
n
t theory.
(4) use
the custome
r
cla
s
sificatio
n
alg
o
rithm a
bove
-
mentio
ned,
and the
tradit
i
onal BP
neural network algorith
m
to cla
ssify cu
stomer.
5.2. Experimental Results
Experimental
data com
e
fro
m
the custom
er
datab
ase registe
r
ed by certai
n E-co
mmerce
enterp
r
i
s
e. Relevant data
of 10041
cu
stomers are ra
ndomly sel
e
ct
ed
from the d
a
taba
se to se
rve
as the
ba
sis for data
mining, an
d ab
stra
ct t
he
re
quire
d custo
m
er valu
e d
a
ta to evalu
a
te
according
to t
he d
e
sig
ned
evaluation i
n
dicato
r
sy
ste
m
. Accordi
n
g
to custom
ers’ overt valu
e
and
potential valu
e, test clu
s
te
rs custo
m
ers i
n
to gol
d
en
cu
stome
r
, silve
r
cu
stome
r
, co
pper cu
stom
e
r
,
gene
ral custo
m
er an
d igno
rable
cu
stom
er; eval
uation
result
s are a
s
sh
own in Table 1.
Table 1. Cu
st
omer
Cla
ssifi
cation
Re
sult
s of Some Website
Customer T
y
pe
Number of
Custo
m
ers
Percentage
%
Profit Contributio
n Proportion
Gold Custome
r
s
786
7.83
53.13
Silver Customers
1278
12.73
30.92
Copper Custome
r
s
2622
26.11
13.94
Gene
ral Custom
ers
3456
34.42
6.13
Negligible Customers
1899
18.91
-4.12
Total 10041
100.00
100.00
From T
able 1
,
we ca
n see
that Gold Custom
e
r
s o
c
cupie
s
7.83%
of the total custome
r
while the p
r
ofits from th
em occu
pie
s
53.13%
of the total. So the gold
cu
stomers play
a
n
importa
nt rol
e
for the
en
terpri
se
s
and
they b
e
tre
a
ted
with
sp
ecial
servi
c
e.
Ho
weve
r, t
h
e
negligibl
e
cu
stomers acco
u
n
t for 18.91%
, who
ma
ke
minus p
r
ofits
for the enterp
r
ise.
In orde
r to te
st the adva
n
tage
s an
d disadvant
ag
es
o
f
improved
al
gorithm i
n
thi
s
pa
per
the improved
model, the traditional BP neural network algo
rithm [7] and K-me
ans alg
o
rithm
[4]
is re
alized in
th pape
r, and
spe
c
ific exp
e
r
imental
re
sul
t
s of all these
three al
gorith
m
s is
sh
own in
Table 2. Fro
m
Table 2 we can se
e that the al
gorithm pre
s
e
n
ted in this pa
per ha
s high
e
r
cla
ssifi
cation
accura
cy than that of th
e ordin
a
ry BP neural net
work alg
o
rith
m and K-me
an
s
algorith
m
.
Table 2. Cla
ssifi
cation Pe
rforma
nce Co
mpari
s
o
n
of different Algori
t
hms
Algorithm
Algorithm in This Paper
Ordina
r
y
BP
Ne
ural
Net
w
ork Algorith
m
K-means Algorith
m
Accuracy
Rate
99.23
%
93.07%
84.36%
Time Consuming
(
S)
18
503
17
6. Conclusi
on
The research
of neural net
work in th
eory
and appli
c
a
t
ion is still d
e
v
eloping. And
how to
corre
c
tly and
effectively carry out
correct and
rea
s
onabl
e cla
s
si
fication on n
e
twork tra
n
sa
ction
cu
stome
r
s, reform net
work marketin
g mode an
d improve custo
m
er man
age
ment and se
rvice
level is also a key to increase the co
mpetitiv
eness of E-comme
rce
enterpri
s
es. Thi
s
pap
er, on
accou
n
t of the sho
r
tag
e
of BP neural n
e
twork in
d
a
ta mining, put
s forward a n
e
w Fo
urie
r b
a
si
s
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No
. 11, Novemb
er 201
3: 643
4 – 6440
6440
neural netwo
rk mo
del to classify netwo
rk tra
de
cust
omers. Experimental
re
sult
s sh
ow that the
improve
d
ne
twork trad
e cla
ssifi
cation
algorit
hm h
a
s enh
an
ced
the a
ccu
ra
cy
of
cust
o
m
er
cla
ssifi
cation,
more reason
able in cl
assif
i
cation results.
Ackn
o
w
l
e
dg
ement
This work is
suppo
rted by sc
ientific rese
arch proje
c
t of the educa
t
ion depa
rtm
ent o
f
Jian
gxi provin
ce (No. G
J
J1
3300
) and 5
2
nd Chi
n
e
s
e p
o
stdo
ctoral f
und (No. 2012
M5212
84
).
Referen
ces
[1]
Ar
w
a
M, Safi
A. CRM Sc
orecar
d - CR
M Performanc
e Meas
urem
e
n
t.
In
te
rn
a
t
i
ona
l
Jou
r
na
l
of
Netw
orked Co
mp
utin
g an
d Advanc
ed Infor
m
ati
on Ma
nag
ement
. 201
2; 2
(
1): 8-21.
[2]
Den
g
WB, He
MS. B2C
Cu
stomer Cl
assifi
cation
Alg
o
rith
m Base
d o
n
Based
on
3D
M.
Jo
u
r
na
l
of
Cho
ngq
in
g Un
i
v
ersity of Posts and T
e
l
e
co
mmu
n
ic
ations (N
atural Sci
enc
e
Editio
n)
. 201
2; 24(4): 56
8-
572.
[3]
Rieh
a
r SF
, N
o
rma T
D
. Stud
y of C
l
assif
y
ing
Custom
ers Method
in
C
R
M.
Journ
a
l of
Co
mp
ute
r
Simulation
. 20
11; 28(8): 2
49-
254.
[4]
Sulma
LE, Stalk E. Res
ear
ch on
Custo
m
er
Class
ifica
t
ion of E-C
o
mmerce W
ebs
ite Base
d o
n
Customer Va
lu
e Anal
ys
is.
Jou
r
nal of Co
mput
er Engi
ne
erin
g
. 2011; 2
4
(6): 2
89-2
94.
[5]
T
eece DJ, Stalk E. Appl
ica
t
ion of Comm
erci
al B
ank C
u
stomer Su
bdi
vision B
a
se
d on K-mea
n
s
.
Journ
a
l of Infor
m
ati
on Ma
nag
ement
. 201
1; 7
(
3): 199-2
09.
[6]
Mardi
y
o
no, Re
ni S, Azlan A.
Intellig
ent Mo
nitori
ng S
y
ste
m
on Pr
edictio
n of Buildi
ng
Dama
ge Inde
x
Using
N
eura
l
-
N
et
w
o
rk.
T
E
LK
OMNIKA Indon
esia
n Jo
urn
a
l
of Electric
al E
n
gin
eeri
ng.
201
2; 10(
1): 1
5
5
-
164.
[7]
Budi
R, Su
pri
y
adi. E
a
rl
y Mo
d
e
l of T
r
affic Si
gn R
e
mi
nder
Based
on
Ne
u
r
al N
e
t
w
ork.
TE
L
K
O
M
N
I
K
A
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2012; 1
0
(2): 7
13-7
22.
[8]
Post LP, Chiris
topher C.
An
Evalu
a
tion of
Methods for Pr
oduc
ing C
R
T
Monitors.
Co
lo
r Research a
n
d
Its Applicatio
n
. 200
7; 14(4): 17
2-18
6.
[9]
Roy
SK, Mark
E.
CRT
Colori
metr
y
B
a
se
d o
n
Improve
d
BP
Neur
al
Net
w
or
k.
Internatio
nal
Revi
ew
o
n
Co
mp
uters an
d Softw
are
. 2009; 38(4): 2
99-
313.
[10]
Li JJ, Rui L
.
Construction
Equipm
ent C
ontro
l Rese
arch Base
d on Pred
ictive
T
e
chnol
og
y.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
n
g
.
2012; 1
0
(5): 9
60-9
67.
Evaluation Warning : The document was created with Spire.PDF for Python.