An Application of Neural Network to Service Quality

bannerclubΤεχνίτη Νοημοσύνη και Ρομποτική

20 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

63 εμφανίσεις

An Appli
cation of Neural Network to Serv
ice Quality

King
-
Jang Yang
(
楊錦章
)
1
, Chin
-
Chow Yang
(
楊錦洲
)
2

and
Bai
-
Sheng Chen
(
陳百

)
2

1
Department of Applied Mathematics, Chung Hua University

No. 707, Sec. 2, Wufu Rd., Hsinchu, 30012, Taiwan, R.O.C.

Tel: 03
-
518
-
63
88
, Fax
: 03
-
518
-
6435

Email:
kingjang@chu.edu.tw

2
Department of Industrial Engineering, Chung Yuan Christian University

No. 200, Chung Pei Rd., Chung Li,
32023,

Taiwan, R.O.C.

Tel:
03
-
265
-
4405
, Fax: 03
-
265
-
4499

Email:
chinchow@cycu.edu.tw


Abstract

In this paper,
we p
resent a classification model of

neural network to measure the performance of service
quality for

a
system

and service

certification provider.

In order to demonstrate the validity of our model, we
also use the case study to build a neural
network mode
l using the backpropagation learning algorithm, and
compare its classification performance against the linear discriminant analysis
.
T
he result shows that
backpropagation neural network model is superior to linear discriminant analysis
model.


Keywords
: ne
ural network, service quality, linear d
iscri
minant

a
nalysis.


1.
I
ntroduction

When many organizations try to find the

ways to
increase the customer
s


satisfaction and loyalty
,

service quality management has become one of the
most critical aspects in organ
izational control.

As a
result, managers have faced deep pressure to measure
the performance of s
e
rvice quality. However
, service
quality is an abstract and elusive thing in construct
because of its three particular features in service
delivery, including

intangibility, heterogeneity and
inseparability.

Hence, it is difficult to assess the
service quality objectively. Basically, th
e service
quality measurement
depend
s

on the customers’
perceptions of the performance on service process.

There always exist
s some gap or difference between
the customers’ expectation
s

and their
actual

perceptions in ser
v
ice quality. The key point is to
understand the importance of various quality
attribu
tes for customers as they assess the quality of
service.

In this study, w
e find out the perspective

of
importance and
satisfaction

using questionnaire
survey to represent the expectation
s

and perception
s

respectively for customers

in performance of service
quality.

This research construct
s

a measurement model
of importance and
satisfaction in service quality from
the view of customers.

Moreover, how to utilize the
data mining technology to disco
ver the business
problems is an
other topic we concern.

Using the
technique of neural network to solve these problem
s
is more popular

t
han traditional methods
.

The
main

difference between neural networks and other
statistical methods is that neural networks have no

assumptions about the statistic
al distribution or

properties of the data
, and therefore tend to be more

useful in practical
situations [5]
.

The classification
characteristic of neural network was used

to

be the
data mining tool to fulfill
the measurement model of
service
quality.

In this study, we use a
practical

exa
mple
of

a

system

and

service

certification
provider

to
demon
strate the advantage of neural
network approach for service qual
ity meas
urement.

First of

all
, we
integrate some important
attributes

of service quality from the viewpoint of
customers t
o understand their expectations
.


The
objective of first stage is

to e
xtract

several

c
ritical
criteria of

service

quality.

Overall survey is

the

second stage. According to the results of first stage,
29 c
riteria are divided into

six dominan
t dimensions
by using fac
tor analysis.

Accordingly
, these six
factors fed as the inp
ut variables

to build a
backpr
opagati
on neural
network (BPN
) model
.

To
evaluate the effectiveness and effi
ciency of the
proposal neural ne
tw
ork classification model
, it is
comp
ared with

linear discriminant analysis (LDA
)
with

respect to six key factors.

The result shows
that the classification performance of neural network
can appropriately discriminate customer categories
on the basis of those six factors.

Besides, on the
accuracy of classification
,

the BPN mo
d
el is superior
to LDA model
.

Finally

on

th
e considerations of
balancing customers’ perceptions and expectations,

the neural network model can provide information to
take decision for improvement.


2.
L
iterature Review

Service quality is often diffi
cu
lt to assess
because of its subj
ective nature
fr
om customers
.

The
measurement model of the service

quality constru
ct
has been domina
ted

by the use of the SERVQUAL
scale

develop
ed by Parasuraman et al. (1988
)
.

This
measur
ement model proposes a five gap
-
based
com
parison between the expectation
s

and
perfo
rmance perceptions of customers
.

For

more
information on SERVQUAL ap
proaches please refer
to [1]. More recent applic
ation
s

of SERVQUAL can
be found

in

[4,

8].

In order to impro
ve and maintain the
relationship

between organization and customer
,

ho
w
to disc
ove
r

the special patterns o
r attributes from the
customers’ database

is

the key
, and
data mining plays
the

imp
ortant role in c
ustomer relationship
management
.
Data mining, also known as

“knowledge discovery in database”
,

is

the process of
discover
ing

meani
ngful pattern
from
customer
databases that
is useful for
decision making
.


Berry
and Linoff

[7]

define

data mining
as
the exploration
and analysis
, by a
utomatic or semiautomatic means
,
of large quantities of data in ord
er to discover
meaningful pattern
s

an
d rules.

Specifi
c techniques
of data mining in
practical
ap
plications include
marker basket an
alysis
, memory based reasoning
,
cluster detection
,
link analysis, decision tree, rule
induction, neural network, and genetic algorithms, etc.
[9]
.

I
n these metho
ds, neural network

approach
es

are
becoming
increasingly popular

in business
.

A
survey of journal articles on business applications
published between
1
988 and 1995 indicates that an
increasing amoun
t of neural network

research is

being conducted for a dive
rse ra
nge of business
activities [2]
.

Many organizations are

going to find
solution
s

f
or

business problems using the techniques
of neural network and data mining, those task
s

are
typically being the
doma
in of the operations
researcher, like forecasting, m
odeling, cl
ust
ering, and
classification [5]
.

According to the widely
applications of neural net
work
, we present a
classification model to evaluate the p
erformance of
service quality
for

a system

and
service

certification
provider
.

We furthermore

buil
d

a
similar mode
l
using the backpropagation learning algorithm and
comp
are its classification performance ag
ainst the
linear dscriminant analysis
.


3.
R
esearch

M
ethodology

There are five major steps for the proposed
neural network approach as shown in Figure 1
.


In
the following section, we use the sample data set of
surveillance audit proc
ess for a system and service

cer
ti
fication provider as the empirical illustration.



Figure 1

Implementa
tion process of proposed model


(1)

Interview

of external and internal
customers:
How the customers fee
l on the important quality
attributes is to be considered as the dimensions of the
auditors


capability, organizational policies,
administration, and certificate issuance when the
certification comp
any
processes surveillance audit.
During the

research period,
we

had interviewed
twenty important customers, and had conducted two
panel discussions by using nominal grou
p

tec
hno
logy
with the front
-
line servers, in

order to

determine the
important quality
attributes and to d
e
sign the valid
questionnai
re .


(2)

Critical to service
quality
attribut
es: W
e

selected the variable which were identified to have
more than 80 % of
significant influenc
e on servic
e

performance th
rough
the pretest ques
tionnaire, and
revised

the
unclear

and uncertain

expressions of
measure
ment

items.


(3)

I
mportance and satisfaction survey
: In this

step,
we

design a concise multiple item scales that contain
29 pairs of Likert
-
type items, where each item is

re
cast into two statements.

O
ne half of

these items
are intended to
measure
the
important degree of
customers’ expect
at
ions about the service categories
being investigated, and
th
e
o
ther 27 matching items
are intended to measure the satisfactory degree of
their

preceptions about the
service per
formance.

The it
ems are presented in a five
-
point response
format with the degrees from strongly important
(sa
tisfactory
) to
strongly unimportant
(dissatisfactory
).


S
ervice

quality is measured by

calculating the ”differen
ce
scores
” between

these

correspo
n
ding

items, that
is

t
he difference between
customers’ expe
c
tative im
p
ortanc
e and perceptive
satisfaction o
n

the quality
attributes.


(4)

Factor analysis
:

I
n order to reduce the
nu
m
bers of i
tems (variables
) , we used th
e

princi
p
al
component an
alysis of princi
pal axes method

with the
varimax criterion of orthogonal rotation to perform a
fact
or analysis
.


Besides, there
is

no theoretical
method to determi
ne the best
-
input variables of

th
e
designed neural network model. Hence
, this
procedure can be perform
ed

as a

generally method to
determine the number of a good subset of inpu
t
variables
.


In this procedure
, we select the variables

which
factor loading

is

greater than

0.45
, and

discover linear
ly related variables and regrou
p

them
into a compound factor
.

These c
ompound factors

will be the input nodes in the next step

of neural
network modeling .


(5)

Neural network modeling: A supervised
learning
algori
thm of backpropagation
is

utilized to
establ
ish the neural network modeling
. A normal
back
p
ropagation neural
(BPN
) m
odel consists of an
in
p
ut layer, one or more hidden layers, and an out
p
ut
layer
. There are two parameters
including
lea
rn
ing
rate

and momentu
m

(
)

required
to define by
the
user
. The theoretical results
showed
that one hidden
layer is s
ufficient for

a

BP network to

a
p
proximat
e

any conti
nu
ous mapping
from

the
input
patterns to the
output
patterns to an arbitrary
degree
freedom
[6]
.


The selections and combinations
of
lea
rn
ing
rate, moment
u
m and the nodes of

hidden
layers primarily affect

the classification performance
.


4.
Empirical

Illustration

4.1
Purpose of study
: The objective of

this

study is
to classify the performance of service quality on the
considerations of auditors


capability
, organizational
pol
icies, administration,

and ce
rtificate issuance for a
system and service

certification provider located in
Taipei, Taiwan.
W
e furthermore make the comparison
of classification performance between the BPN
model and LDA model.


4.2
Sample:

The data for this
study were collected
from 529 firms th
at responded to a mailed survey.

The responding fi
rms were primarily involved
110

firms of High
-
tech industry, 5
1

firms of traditional
manufact
u
ring
, 54

firms of construction industry,
11
6
f
irms of machinery and equipm
ent
, 73 firms of
chemical and plastic products, 88 firms of service
industry, 26 firms of others and
11

firms missed
.

The

response rate based on 529 ret
urn
s was 25.68 %.

There were 467 valid sam
p
les to perform the
following a
n
a
l
ysis
,

and the valid rate o
f samples
w
as
22.67%.

For the reliability test, the alpha value
s

of
importance
, satisfaction and difference scores
evaluation are
all
greater than

0.92.


4.3
Factor anal
y
sis:

In this step, the “difference
scores” was used to perform the factor analysis wi
th
the principal com
p
onent e
stimation

method and the
varimax rotation with Kaiser normalization.

Furthermore, the original set of 29

variables was
reduced to 6 principal compou
n
ded factors, where
th
e
55.
25% of variance was explained. The regrou
p
ing
measur
ement variables can be set into six different
dimensions included responsiveness
(
)
, assurance
(
)
, reliability

(
)
, empathy
(
)
, value
-
added
service

(
)
,

and tangibility
(
)
inc
l
uded

7,5,5,4,3

and

3 items respectively. The next step was to assess
the performance of the services provider, which was
to be
explained by these
six critical factors
.

Also
,
t
hese six factors were used to be the input variables in
the next step of neural network modeling .


4.4
Neural network modeling:

The 467 respondent
firms were randomly separated into two groups,
namely
, 75
% for training patterns and 25% for testing
pattern
s.


The three parameters of learning rate,
momentum and the
number

of nodes in the hidden
layer should be defined for back
p
ropagation network
modeling. In the training model, the six factors were
fed as input nodes as discussed in section 4
.
3.

For
the out
p
ut nodes decision, we then divided the
sampl
e into three categories of high
, medium and low
service quality for services provider’
s

performance
based on customers’ judgments.


W
e adopted the
range of 0.6
-
0.9 and 0.1
-
0.4 to be the decisions of
learning rat
e and momentum [3].

A very rough
rule
-
of
-
thumb for number

of

hidden nodes defined as

, where
,
,

and

represent the
number of train
ing patterns
,
ou
tput nodes and input
nodes respective
l
y.

Then

the root
-
mean
-
square

error
(RMSE)

and classification
rate are the measurement
indictors to validate the performance of the training
model.
T
hrough several trial
-
and
-
error experiments,
the structure of 6
-
9
-
3 mod
el had the best performance.
F
urthermore, five nodes for the second hidden layer
had validated to be the highest classification rate and
lowest RMSE.
I
n order to obtain the optimal
combination of learning rate and momentum, it
then
used sixteen combination
s settings as: (0.6, 0.1), (0.6
,
0.2), (0.6, 0.3), (0.6, 0.4),

,

(0.9, 0.4)
.

In term
s of
learning rate and momentum, the best setup is (0.6,
0.3) on the considerations of
classification

rate and
RMSE.

T
he relationship between the RMSE and
the numbers of
learning iterations for the selected
6
-
9
-
5
-
3 model is in Figure 2.





Figure 2

RMSE versus numbers of learning
i
terations


4.5

N
eural
network

and linear discriminant
analysis:

To validate the efficiency of the proposed
BPN mode
l for the practical application,
it is

compared with linear discriminant analysis (LDA)
with respect to six critical factors to obtain a linear
model.

T
his model is a linear combination of
responsiveness, assurance, reliability, empathy,
value
-
added servi
ce and tangibility factors to separate
the service quality into three groups.

T
he linear
discriminant fu
n
ctions by using stepwise regression
method for the high
(
)
, medium
(
)
and low
(
)
service quality are as following:


The thresho
l
d value is
, meaning that if
the discriminant function

, then t
he
service level

on

the service performance of the
customer perceived is ranked to be high.
I
f
, it is ranked to be medium, and if
, it is rank
ed

to be low.

T
he LDA
model c
lassifies

correctly in 71.73%

of the sample.
T
he classification rate of low service quality is only
43.80%, it is clear that the LDA model is not good
enough for classification.

For the best BPN model,
the classification rates of training and testing are
94.40% and 87.71%

separately
.


And the RMSE
s

of
training and testing are
0.1174 and 0.1516

separately
.
T
he BPN model has
85.86% network accuracy in
class
ifying

service quality categories.


F
rom the
classification results of
accuracy
, the BPN model is
superior to LDA model on all three

categories and
total

accuracy is

shown in Table 1.

T
herefore, the
backpropagation neural network model has been
demonstrated as a good
method
because

it

predict
well in classification problem.


Table 1

Classification results of the two models


BPN

LDA

H
igh se
r
vice quality

88.09%

66.67%

Medium service quality

84.21%

82.31%

Low service quality

85.29%

43.80%

Total accuracy

85.86%

71.73%


Choosi
ng one specific improvement strategy
over another is
difficult for managers.

B
ut it is also
the obligation for

managers to do

that
, since
improvement strategy can make the allocation of
resources more effective
under
organizational limited
resources
.


T
o obtain

further improve
ment
information, we suggest using
statistical

analysis to
extract the data from the low
service category.

A

study
of this category should provide some insight
for managers to make decision about improvement
strategy.


5.
C
on
c
lusion

The main result of this research are twofold.

The first is that we illustrate the procedures of service
qualit
y measurement from the customers’ perspective.

Our proposal model could be used to predict service
quality classifications by applying neural network
r
easonably well.

The

second result shows that the
classification performance of backpr
opagation neural
n
etwork model
is

better th
an linear discriminant
analysis
.

The
backpropagation neural

network is

an
effective classification
method to deal with the high
vo
l
umes of uncer
tain information from customers


subjective
perceptions
. For further study
, the geneti
c
algorithm
can

be

used to constr
u
ct optimal
architect
u
res and parameters of networks
to improve
the decision quality

in applications
.


References

[1]

A
. Paras
uraman, V
. A
. Zeitham
l and L. L.
Berry, “SERVQUAL:

a multiple
-
item scale for
measuring

consumer

perce
ptions of service
quality
,”
Journal of Retailing
, vol. 64, pp12
-
40
,
Spring l988 .

[2]

B. K. Wong, T. A. Bodnovich and Y. Selvi,
“N
eural n
etwork applications in business
: a
review and analysis of
th
e lit
e
rature
(
1988
-
95
)
,”

Decision
Support Systems
, vol. 19,
pp3
01
-
320, 1997
.

[3]

C. T. Su, T. Yang and C. M. Ke, “
A neural
network a
p
proach for semiconductor

wafer
post
-
sawing inspection,”

IEEE Trans.

o
n
Semicond
u
ctor M
anufacturing, vol. 15, no. 2,
May 2002
.

[4]

H. Kang and G
.

Bradley, “
Measur
ing the
performance of IT servi
c
e
s: an assessment of
SERVQUAL,”

In
ternational J
o
urnal of
Account
ing Infor
mation
Syst
ems, vol. 3
, pp
151

-
164, 2002
.

[5]

K. A. Smith, J. N. Gupta,

Neural networks in
business: techniques and applications for the
operations
research
,


Computers & Operations
Resea
rch,
vol.

27, pp
1023
-
1044, 2000.

[6]

L. Fausett,

Fu
ndamentals of Neural

Ne
tworks:
Architecture
, Algorithms and
App
lications,

New Jersey: Prentice
-
Hall, 1994
.

[7]

M. J
. Berry

and G
. Linoff
, Data Mining
Techniques

for

Mark
eting
, Sales and Customer
Supp
ort
, New York
: Wiley
,
1997
.

[8]

M. K. Bra
dy,
J. J. Cronin and R. R. Brand,
“Performance
-
only measurement of service

quality
: a replication and extensio
n,


J
o
u
rna
l of
Business Research, vol. 55
, ppl7
-
31
, 2002
.

[9]

T. Hong and
I
. Han, “Knowledge
-
based data

mining of news informa
tion on the Internet
using cog
nitive maps and neural networks,


Exp
ert

Systems

with
app
lications, vol. 23,
ppl
-
8
, 2
002
.

[10]

C. C. Yang,

I
m
provement actions based on the
customers


satisfac
tion
survey
,


TQM &
Business Excellence, vol.

14, no. 8, pp919
-
930,
200
3.