You are on page 1of 6

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/224300451

A new weighting method for detecting outliers in IPA based on Choquet


integral

Conference Paper · January 2008


DOI: 10.1109/IEEM.2007.4419545 · Source: IEEE Xplore

CITATIONS READS

10 187

4 authors, including:

Hsiang-Chuan Liu Der-bang Wu


Asia University National Taichung University of Education
126 PUBLICATIONS   933 CITATIONS    47 PUBLICATIONS   232 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Hsiang-Chuan Liu on 20 November 2015.

The user has requested enhancement of the downloaded file.


A New Weighting Method for Detecting Outliers in IPA Based on Choquet
Integral

H.-C. Liu1, C.-C. Chen2, 3, D.-B. Wu2, 4, Y.-D. Jheng2


1
Department of Bioinformatics, Asia University, Taichung, Taiwan
2
Graduate Institute of Educational Measurement and Statistics, Taichung University, Taichung, Taiwan
3
Department of General Education, Min-Hwei College of Health Care Management, Tainan, Taiwan
4
Department of Mathematics Education, Taichung University, Taichung, Taiwan
II. IMPORTANCE-PERFORMANCE ANALYSIS
Abstract - When interactions among items of survey
exist in Importance-Performance Analysis (IPA), the Importance-performance analysis (IPA) was
traditional equal weighting method for detecting outliers in introduced to study customer satisfaction as a function of
IPA is not always available. In this paper, we suggest to use both expectation related to importance and performance.
Choquet integral based on two fuzzy measures, the well [1, 2]. IPA involves plotting mean ratings on each of the
known fuzzy measures, λ-measure, proposed by Sugeno, and
γ-measure, proposed by our previous study, to improve this
items from the survey results in a two-dimensional grid to
situation. A real data experiment by comparing the numbers produce a four-quadrant matrix that identifies areas of
of reduced outliers and the changing numbers of items needing improvement as well as areas of effective
which were needed improvement was conducted. The performance [5, 6]. The typical four-quadrant matrix in
performances of traditional equal weighting method, λ- IPA was shown in Figure 1.
measure Choquet integral weighting method, and γ-measure
Choquet integral weighting method for detecting outliers in
IPA were compared. Experimental result shows that the γ- (II) (I)
measure Choquet integral weighting method outperforms
the other two weighting methods. Keep up
Importance

Concentrate here The good work


Keywords - IPA, outliers, λ-measure; γ-measure,
Choquet integral. (III) (IV)

Low priority Possible overkill


I. INTRODUCTION

When interactions among items of survey exist in performance


Importance-Performance Analysis (IPA), [1, 2], the
traditional equal weighting method for detecting outliers Fig. 1. Four-quadrant matrix in IPA.
in IPA is not always available. In this paper, we suggest to
use Choquet integral based on two fuzzy measures, the Let I j , j = 1, 2,..., n denote item j of the survey,
well known fuzzy measures, λ-measure [3] and γ-measure, let xij , yij ∈ {1, 2, 3, 4, 5} , i = 1, 2,..., N ; j = 1, 2,..., n denote
proposed by our previous study [4] to improve this
situation. A real data experiment by comparing the performance ratings and importance ratings of item j , by
numbers of reduced outliers and the changing numbers of consumer i , respectively.
items which were needed improvement was conducted. 1 N
1 N

The performances of traditional equal weighting method x. j =


N
∑x
i =1
ij , y. j =
N
∑y
i =1
ij , j = 1, 2,..., n . (1)
(E-method), λ-measure Choquet integral weighting
method (λ-method), and γ-measure Choquet integral
1 n 1 n
weighting method (γ-method) for detecting outliers in IPA x= ∑
n j =1
x. j , y = ∑ y. j
n j =1
(2)
were compared. Experimental result shows that the γ-
method outperforms the other two methods.
This paper was organized as followings. The By IPA, the strengths and weakness can be easily
Importance-Performance Analysis (IPA) was introduced identified. If x. j < x , y. j > y , then ( x. j , y. j ) is into
in section II and fuzzy measures were reviewed in section
quadrant II. The item j with high importance but low
III. Our new fuzzy measure, γ-measure, was described in
section. IV. The weighting method of Choquet integral performance should be placed in high priorities to
was described in section V. Results and discussion were improve service quality. On the contrary, If x. j > x , y. j < y ,
described in section VI and final section was the then ( x. j , y. j ) is into quadrant IV. The item i with high
conclusions.
performance but low importance shows that the
consumers do not appreciate the efforts. The resources
might be considered to be waste.
Let xi . = 1 ∑ xij , yi. = 1 ∑ yij , i = 1, 2,..., N ( )
n n
(ii) ∏ ⎡1 + λ s ( xi )⎤ = λ + 1 > 0, s ( xi ) = gλ { xi }
n
(3) (9)
n n ⎣ ⎦
j =1 j =1 i =1
qH ( x ) = q3 ( x ) + 1.5 ⎡⎣ q3 ( x ) − q1 ( x ) ⎤⎦ ,
qL ( x ) = q1 ( x ) − 1.5 ⎣⎡ q3 ( x ) − q1 ( x ) ⎦⎤ (4)
IV. NEW FUZZY γ-MEASURE [4]
Where q3 ( x ) , q1 ( x ) are the third quartile and the first
quartile of the data xi . , i = 1, 2,..., N , respectively.
γ-measure μ on a finite random set X ={ X1, X2,..., Xn}
If xk . ∉ ⎡⎣ qL ( x ) , qH ( x ) ⎤⎦ then xk . is an outlier of
is a set function γ : 2X →[0,1] satisfying the following
performance ratings { x1. , x2. ,..., xN . } . Similarly, we can
conditions:
define the outlier of importance ratings { y1. , y2 ,..., yN } .
(i) Xj , i =1,2,..., N, J =1,2,..., n , where Xj = xj are ratings of
i i i

The ratings of consumer with unfair outlier should be


1n i
deleted from the data. strategy j by student i , Yi = ∑Xj , i =1,2,..., N
n j=1
(10)

III. FUZZY MEASURES


i
{ i i
}
(ii) X(1) , X( 2) ,...X( n) , i =1,2,..., N are the permutations of

The equation (3) shows that the mean ratings with { X , X ,...X } , i =1,2,..., N , respectively,
i
1
i
2
i
n

equal weighting
1
n
. When interactions among items of
i
{
satisfying x(1) ≤ x( 2) ≤,... ≤ x( n) , i =1,2,..., N
i i
} (11)

survey exist in IPA, the traditional equal weighting


method for detecting outliers in IPA is not always
{
A(i j) = X(i j) , X(i j+1) ,..., x(in) , i =1,2,..., N } (12)

available. In this paper, we consider the unequal (iii)


1+ Ry.xj
weighting of Choquet integral based on a new fuzzy
measure, γ-measure, proposed by our previous study [4].
γ (φ) = 0,γ ( X) =1,γ { Xj } =
2
, j =1,2,..., n ( )
(13)

Ry.xj , j =1,2,..., n are the simple correlations of


A. Fuzzy Measure
dependent variable Y on random
A fuzzy measure μ on a finite set X is a set function
variable Xj , j =1,2,..., n
gμ : 2 →[ 0,1] satisfying the following axioms:
X

RY.Ai = Ry. Xi , j =1,2,..., n are the multiple


( j) ( i i
)
( j) , X( j+1) ,..., X( n) .
(i) gμ (φ) = 0, gμ ( X ) =1 (boundary conditions) (5) correlations of dependent variable Y on random

(ii) A ⊆ B ⊆ X ⇒ gμ ( A) ≤ gμ ( B) (monotonicity) (6)


i
(
vector X( j) , X( j+1) ,..., X( n) . , j =1,2,..., n
i i
)
1+ Ry.Ai

B. Singleton Measure [7]


( )
γ A(i j) =
2
( j)
, j =1,2,..., n,i =1,2,..., N (14)

The singleton measure s of a fuzzy measure μ on a


finite set X is a function s : X → [ 0,1] satisfying V. CHOQUET INTEGRAL [1, 3, 8, 9, 10]

s ( x ) = μ ({ x} ) , ∀ x ∈ X (7) Let μ be a fuzzy measure on a finite set X. The


Choquet integral of f : X → R+ with respect to μ is denoted
by
C. λ -measure [3]
∑ ⎡⎢⎣ f ( x( ) ) − f ( x( ) )⎤⎥⎦ g μ ( A( ) )
n

For given singleton measure s, let λ ∈ ( −1, ∞) , a λ - ∫ C fd μ =


j =1
j j −1 j (10)

measure, gλ , is a fuzzy measure on a finite set X, X = n , where ( )


f x( 0 ) = 0 , ( )
f x( j ) indicates that the indices have
satisfying been permuted so that
(i) A, B ∈ 2 X , A ∩ B = φ , A ∪ B ≠ X ( ) ( )
0 ≤ f x(1) ≤ f x( 2 ) ≤ ... ≤ f x( n ) ( ) (11)

⇒ gλ ( A∪ B) = gλ ( A) + gλ ( B) + λgλ ( A) gλ ( B) (8) {
A( i ) = x( i ) , x( i +1) ,..., x( n ) }
Step4. Interpret the matrix of the importance-performance
analysis.

D. Results
VI. RESULTS AND DISSCUSIONS Five outliers of the equal weighting and λ-Method
were deleted. Nine outliers of the γ-Method were removed.
A. Why IPA? These data demonstrated that the γ-Method is better than
In the present, there are overall 146 universities and those of the E-Method and λ-Method. The numbers of the
16 junior colleges in Taiwan. Previously, university deleted outliers were shown in Table I.
management and teacher-selves all concentrated on
research work, because of teachers promoting based on
evaluating their research. Recently, administrator of
education and manager of university have a change to TABLE I
concentrate on teaching quality. DELETED OUTLIERS OF EACH METHOD
It is important to understand what teaching Methods E-Method λ-Method γ-Method
evaluations a teacher will develop or improve in high
Outliers 5 5 9
priority, and what in low priority? Importance-
performance analysis (IPA) offers a simple, yet useful
method for simultaneously considering both the
importance-performance dimensions when evaluating or
defining evaluations. This study applies IPA to evaluate The coordinates of all 129 data with outliers were
teaching evaluations in a junior college in Taiwan. The shown in Table II. The origin is (3.8630, 3.8594).
survey data of evaluation of a mathematics teacher were
as an example used in this study.
TABLE II
COORDINATE OF ALL 129 DATA WITH OUTLIERS
B. Participates Question Importance Performance
In this study, a teaching satisfaction questionnaire of
1 3.8837 3.8372
15 items was surveyed and 129 college students were
asked to respond. The performance for each strategy in 2 3.9070 3.8682
the survey can be computed by the mean value, while the 3 3.8295 3.8682
importance for each strategy was evaluated by E-method, 4 3.8527 3.8372
λ-method, and γ-method, respectively. 5 3.7984 3.8062
Then, a four-quadrant importance-performance map 6 3.7364 3.8450
can be used to offer the teacher and his manager how they
7 3.8760 3.8915
develop and improve their teaching evaluations.
8 3.8527 3.8837
C. Steps of Calculating 9 3.7984 3.8295
To construct the map of the importance-performance 10 3.8450 3.8217
analysis without outliers in this study, a four step 11 3.9612 3.8837
procedure is as follows. 12 3.9147 3.8915
Step1. Three kinds of value of mean performance and 13 3.9225 3.8992
mean importance for each college student are
14 3.8915 3.8760
gained by calculating the mean value,
15 3.8760 3.8527
( xi. , yi. ) , i = 1, 2,...,129 , using E-Method, λ-Method,
Note: The origin is (3.8630, 3.8594).
and γ-Method, respectively.
Step2. After deleting all outliers from the data in Step 1,
Three kinds of modified value of mean performance
and mean importance for each strategy are gained
by calculating the mean value ( x. j , y. j ) , j = 1, 2,...,15 ,
The distributions of all 129 data with outliers were
and ( x , y ) , using E-Method, λ-method, and γ- shown in Fig. 1. There are 2 items, item 3 and 8, to assign
method, respectively. to quadrant 3. It means that these 2 items were with high
Step3. Plot IP map. First, construct a two-dimensional importance but low performance. It should be placed in
grid by plotting a horizontal line as x-axis and a high priorities to improve service quality.
vertical line as y-axis with origin ( x , y ) where.
Then, plot these order-pairs, ( x. j , y. j ) , j = 1, 2,...,15 ,
in a four-quadrant matrix.
Fig. 2. Distributions of all 129 data with outliers. Fig. 3. Distributions of the equal weighting and λ-Method without five
outliers.

By using the equal weighting and λ-Method


respectively, five outliers were removed. The coordinates By using the γ-Method, nine outliers were removed.
of the equal weighting and λ-Method without five outliers The coordinates of the γ-Method without nine outliers
were shown in Table III. The origin is (3.9430, 3.677). were shown in Table IV. The origin is (4.0044, 4.0439).

TABLE IV
COORDINATE OF THE EQUAL WEIGHTING AND
TABLE III γ -MEASURE METHOD WITHOUT NINE OUTLIERS
COORDINATE OF THE EQUAL WEIGHTING AND
λ -MEASURE METHOD WITHOUT FIVE OUTLIERS Question Importance Performance
Question Importance Performance 1 4.0250 4.0167
1 3.9516 3.9516 2 4.0417 4.0500
2 3.9839 3.9758 3 3.9750 4.0417
3 3.9032 3.9839 4 3.9833 4.0167
4 3.9194 3.9435 5 3.9417 3.9917
5 3.8790 3.9194 6 3.8750 4.0333
6 3.8145 3.9516 7 4.0250 4.0833
7 3.9516 4.0000 8 3.9833 4.0750
8 3.9274 3.9919 9 3.9417 4.0167
9 3.8790 3.9355 10 3.9833 4.0000
10 3.9274 3.9194 11 4.1083 4.0833
11 4.0403 4.0000 12 4.0583 4.0750
12 4.0000 3.9919 13 4.0833 4.0917
13 4.0161 4.0081 14 4.0500 4.0583
14 3.9919 3.9839 15 3.9917 4.0250
15 3.9597 3.9597 Note: The origin is (4.0044, 4.0439).
Note: The origin is (3.9430, 3.677).

The distributions of the γ-Method without nine


The distributions of the equal weighting and λ- outliers were shown in Fig. 4. There is only 1 item, item 8,
Method without five outliers were shown in Fig. 3. There to assign to quadrant 3. It means that this item was with
are still 2 items, item 3 and 8, to assign to quadrant 3. It high importance but low performance. This demonstrated
means that these 2 items were with high importance but that the γ-Method is better than both of the equal
low performance. It should be placed in high priorities to weighting and λ-Method.
improve service quality.
REFERENCES

[1] S. R. Magal, & J. C. James, “Importance-performance


analysis,” Journal of Marketing, vol. 41, no. 1, pp. 77-79,
1977.
[2] S. R. Magal, & N. M. Levenburg, “Using importance-
performance analysis to evaluate e-business strategies
among small businesses,” In Proceedings of the 38th
Hawaii International Conference on System Sciences, p.
176, 2005.
[3] M. Sugeno, Theory of fuzzy integrals and its applications,
unpublished doctoral dissertation, Tokyo Institute of
Technology, Tokyo, Japan, 1974.
[4] H.-C. Liu, “The Choquet integral regression model based
on γ -complete measure,” Journal of educational research
Fig. 4. Distributions of the γ-Method without nine outliers. and development. vol. 2, no. 4, pp. 87-107, 2006. (in
Chinese)
[5] L. A. Graf, , M. Hemmasi, & W. Nielsen, “Importance-
VII. CONCLUSION satisfaction analysis: a diagnostic tool for organizational
change,” Leadership and Organization Development
With the use of IPA by using E-Method, λ-Method, Journal, vol. 136, pp. 8-12, 1992.
and γ-Method, this matrix map can help teachers [6] W. Skok, A. Kophamel, & I. Richardson, “Diagnosing
understand deeply the things about his teaching information systems success: importance-performance
maps in the health club industry,” Information &
evaluations in the classroom.
Management, vol. 38, pp. 409-419, 2001.
The evaluations with extremely important but fair [7] H.-C. Liu, W.-C. Lin, W.-S. Weng, “A choquet integral
performance can be identified easily by IPA for regression model based on a new fuzzy measure,” 12th
immediate improvement. Conversely the evaluations with International Conference on Fuzzy Theory & Technology,
slightly important but excel performance are possibly JCIS, Salt Lake City, Utah, U.S.A., Jul., 2007, (accepted, in
overemphasized, which indicates that too much efforts are press).
expended here. [8] G. Choquet, “Theory of capacities,” Annales de l’Institut
While suppose that we focus only on areas of low Fourier, vol. 5, pp. 131-295, 1953.
performance, it may be of little value, if these areas are [9] G. Shafer, A Mathematical Theory of Evidence. Princeton
University Press, Princeton, New Jersey, 1976.
not important. Similarly, an area being important does not
[10] Z. Wang, and G. J. Klir, Fuzzy measure theory. Plenum
mean that efforts should be expended in that area. Press, New York, 1992.
Performance may be considered simultaneously. [11] A. P. Dempster, “Upper and lower probabilities induced by
A comparison of three IPA methods was discussed in multi-valued mapping,” Annals of Mathematical Statistics,
this study. The surveying data demonstrated that the γ- vol. 38, pp. 325-339, 1967.
Method is better than those of E-Method and λ-Method.

View publication stats

You might also like