JID: EOR

ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

European Journal of Operational Research 000 (2014) 1–15

Contents lists available at ScienceDirect

European Journal of Operational Research

journal homepage: www.elsevier.com/locate/ejor

Decision Support

A group decision-making approach to uncertain quality function

deployment based on fuzzy preference relation and fuzzy majority
Hong-Bin Yan∗, Tieju Ma
School of Business, East China University of Science and Technology, Meilong Road 130, Shanghai 200237, P.R. China

a r t i c l e i n f o a b s t r a c t

Article history: Quality function deployment (QFD) is one of the very effective customer-driven quality system tools typically
Received 28 December 2013 applied to fulfill customer needs or requirements (CRs). It is a crucial step in QFD to derive the prioritization
Accepted 15 September 2014
of design requirements (DRs) from CRs for a product. However, effective prioritization of DRs is seriously
Available online xxx
challenged due to two types of uncertainties: human subjective perception and customer heterogeneity. This
Keywords: paper tries to propose a novel two-stage group decision-making approach to simultaneously address the two
Quality management types of uncertainties underlying QFD. The first stage is to determine the fuzzy preference relations of different
Uncertain QFD DRs with respect to each customer based on the order-based semantics of linguistic information. The second
Group decision-making approach stage is to determine the prioritization of DRs by synthesizing all customers’ fuzzy preference relations into
Fuzzy preference relation an overall one by fuzzy majority. Two examples, a Chinese restaurant and a flexible manufacturing system,
Fuzzy majority are used to illustrate the proposed approach. The restaurant example is also used to compare with three
existing approaches. Implementation results show that the proposed approach can eliminate the burden
of quantifying qualitative concepts and model customer heterogeneity and design team’s preference. Due
to its easiness, our approach can reduce the cognitive burden of QFD planning team and give a practical
convenience in QFD planning. Extensions to the proposed approach are also given to address application
contexts involving a wider set of HOQ elements.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction Quality function deployment (QFD) is one of the very effective

Nowadays, with short life-cycles and dynamic competition in
global markets, the major challenge of any product-oriented firm is
how to efficiently design, develop, and manufacture new products
that will be preferred more by customers than those offered by com-
petitors (Chen & Ko, 2010). Essentially, a product’s success depends
largely on how it meets customer needs or requirements (CRs). On
one hand, the CRs are obtained through a survey conducted by the
marketing department, and the output of this is a list of qualitative
customer attributes, such as “easy to use”, “resistant” or “durable”. On
the other hand, the design team has to make the product specifica-
tions satisfy what the customers want. The design specifications are
based on engineering properties with a quantitative nature, such as
“automated guided vehicle”, “storage and retrieval system” or “pro-
grammable logic controller”. In this sense, conflict can arise between
marketing and engineering departments, as they speak different lan-
guages (Hauser & Clausing, 1988).
∗
Corresponding author. Tel.: +86 21 64250013; fax: +86 21 64252015.
E-mail addresses: hongbinyan_0214@126.com (H. Yan), tjma@ecust.edu.cn (T. Ma).
customer-driven quality system tools typically applied to fulfill cus-
tomer needs and, more importantly, to improve customer satisfac-
tion (Chan & Wu, 2002a,b; Chen & Ko, 2011). The most commonly
seen QFD consists of four inter-linked stages: Stage 1 to translate CRs
into design requirements (DRs); Stage 2 to translate important DRs
into product characteristics; Stage 3 to translate important product
characteristics into manufacturing operations; and Stage 4 to trans-
late key manufacturing operations into operations and control. Each
stage’s important outputs (HOWs), generated from the stage’s in-
puts (WHATs), are converted into the next stage as its inputs (new
WHATs), i.e., each stage can be described by a two dimensional matrix
of “WHATs” and “HOWs”, which is easy and convenient to deal with
in practice (Chan & Wu, 2005).
The first stage of QFD, also known as house of quality (HOQ), is
of fundamental and strategic importance, since it is in this stage that
the CRs (WHATs) for the product are identified and converted into ap-
propriate DRs (HOWs) to fulfil customer satisfaction. In other words,
HOQ links the “voice of the customer” to the “voice of the technician”,
through which the process and production plans can be developed in
the other stages of the QFD system, as depicted in Fig. 1. The structures
and analyzing methods of the other three QFD stages are essentially
the same as the first one (Liu & Wu, 2008). Therefore, instead of the

http://dx.doi.org/10.1016/j.ejor.2014.09.017
0377-2217/© 2014 Elsevier B.V. All rights reserved.

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS
2 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
Fig. 1. The house of quality.
entire four stages, most QFD studies focus mainly on the first stage
to prioritize DRs (HOWs) from CRs (WHATs; e.g., Chen et al., 2006;
Ho, Lai & Chang, 1999; Yan, Ma & Li, 2013), which is also the main
focus of our current work.
Successful implementation of QFD often requires a significant
number of subjective judgments from both customers in a targeted
market and a QFD design team in a firm. Such a number of subjec-
tive judgments will inevitably generate the following two types of
uncertainties for prioritizing DRs in the process of QFD planning.
• The first type of uncertainty is the human assessment and judg-
ment on qualitative attributes, which are always subjective and
imprecise. Hence, the input information of human perception can
be ambiguous, which presents a special challenge to effective pri-
oritization of DRs (Chen et al., 2006).
• The second type of uncertainty is the involvement of many cus-
tomers and the QFD design team in the evaluation of input infor-
mation of QFD. Input information may have an uncertainty asso-
ciated with customer heterogeneity because each customer may
have a different opinion (Kwong et al., 2011; Wang, 2012).
In the past two decades, we have witnessed many studies focusing
on these two types of uncertainties in the literature, as reviewed in
Section 2.1.
Regarding the first type of uncertainty, it has been found that
all existing studies of uncertain QFD have quantified subjective
judgments in terms of fuzzy sets (Zadeh, 1965), which is in fact
the process of transforming an ordinal scale into a cardinal scale
that represents an “arbitrary passage”. Such a quantification may
sometimes be dangerous (Han, Kim & Choi, 2004), since the fuzzy set
based semantics of linguistic labels is often defined subjectively and
context-dependently. Consequently, it is easy to generate different
results by choosing different scales from which to draw the ordinals.
Moreover, even if the quantification process used is rational, existing
studies of uncertain QFD perform calculations with the associated
fuzzy membership functions of linguistic labels based on fuzzy
extension principle (Zadeh, 1965). Such a procedure has, as any
fuzzy computation-based approach, an unavoidable limitation of
information loss caused by the process of linguistic approximation,
which consequently implies a lack of precision in the final result and
has been well verified by Herrera and Martínez (2000) in the general
context of fuzzy linguistic decision-making. Finally, as observed by
Tidd and Bodley (2002), many tools even if available within the firm
are not used since they are too complex to be used. Since it is complex
and time-consuming to apply QFD in practice (Chan & Wu, 2005), the
burden of quantifying qualitative concepts and calculation in terms of
fuzzy sets may build an obstacle for the design team to use QFD. With
respect to the second type of uncertainty, due to the heterogeneity
of customer inputs, it is necessary and important to consider the cus-
tomer variability so as to derive a robust prioritization of DRs (HOWs;
Kwong et al., 2011). As a basic element underlying group decision-
making, the concept of fuzzy majority is accepted by most of its
members in practice, since it is quite difficult for the solution to be ac-
cepted by all users (Kacprzyk, Zadrozny, Fedrizzi & Nurmi, 2008). Such
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
[m5G;October 10, 2014;12:11]
a concept of fuzzy majority may represent the design team’s prefer-
ence, which plays an important role in QFD. Therefore, it may provide
a better solution to prioritize the DRs by considering the concept of
fuzzy majority in uncertain QFD, which is missed in the literature.
Toward this end, the main focus of this paper is to simultaneously
cope with the two types of uncertainties underlying QFD by a novel
group decision-making approach. Firstly, the proposed approach per-
forms computations solely based on the order-based semantics of
linguistic input information. Moreover, it performs the group aggre-
gation of fuzzy preference relations based on the concept of fuzzy
majority underlying group decision-making. An exponential Regular
Increasing Monotone (RIM) quantifier is used to express the fuzzy ma-
jority. Finally, it utilizes the fuzzy majority to derive a consensus de-
gree for the uncertain QFD problem. As we shall see, the proposed ap-
proach, on one hand, can eliminate the burden of quantifying qualita-
tive concepts in QFD; on the other hand, it incorporates the fuzzy pref-
erence relation and fuzzy majority into uncertain QFD so as to model
the customer variability and the design team’s attitudinal preference.
The rest of this paper is organized as follows. Section 2 reviews
existing approaches to uncertain QFD and formulates our research
problem. Section 3 presents how to derive the fuzzy preference rela-
tions from users’ subjective judgments. Section 4 proposes the group
aggregation of individual fuzzy preference relations based on the con-
cept of fuzzy majority. A consensus degree based on fuzzy majority
is also given for the uncertain QFD problem. Section 5 examines two
examples, a Chinese restaurant and a flexible manufacturing system,
to show the effectiveness of the proposed approach. The restaurant
example is also used to compare with three existing studies. For the
completeness of this study, some considerations and potential ways
are discussed in Section 6 to handle other factors in a potentially
more comprehensive HOQ model. Finally, the paper is concluded in
Section 7 with some remarks.
2. Literature review and our QFD framework
2.1. Literature review
Several attempts have been made in order to cope with the two
types of uncertainties underlying the QFD problem: human subjec-
tive perception and customer heterogeneity. To cope with the first
type of uncertainty, numerical studies have been conducted on how
to prioritize DRs (HOWs) with fuzzy linguistic variables (Zadeh, 1975)
semantically represented by fuzzy sets (Zadeh, 1965). For example,
Khoo and Ho (1996) developed an approach based on possibility the-
ory and fuzzy arithmetic to address the ambiguity involved in various
relationships and outlined the framework of a fuzzy linguistic QFD.
Zhou (1998) proposed an approach to prioritize DRs through a fuzzy
ranking procedure and to optimize improvements using a mixed in-
teger programming approach, in which the relative importance of
CRs were determined by the analytical hierarchy process (AHP) and
were assumed to be crisp numbers. Wang (1999) used fuzzy arith-
metic to compute the technical importance ratings of DRs and the
outranking approach based on possibility and necessity measures to
prioritize DRs. Shen, Tan & Xie (2001) employed fuzzy arithmetic to
calculate the fuzzy priority weights of DRs and defuzzified them using
the mean of maxima method and the centroid defuzzification method.
Ertay, Kahraman & Ruan (2005) prioritized DRs by taking into account
the degree of interdependence between CRs and DRs, and the inner
dependence among them. Liu (2005) devised a method that could
prioritize DRs without knowing their exact membership functions by
means of fuzzy weighted aggregation (FWA; Liou & Wang, 1992; Kao
& Liu, 2001). Chen et al. (2006) calculated the priority weights of DRs
using the FWA and fuzzy expected value (FEV) operator (Liu & Liu,
2002). Kwong, Chen, Bai & Chan (2007) developed a new methodol-
ogy to calculate the importance weights of DRs based on the fuzzy
expert systems approach. Chen and Ko (2011) considered the four
JID: EOR
ARTICLE IN PRESS
H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
stages of QFD to build a series of FWA approaches for determining the
fulfillment levels of each decision outcome for customer satisfaction.
The readers can refer to Chen and Weng (2006), Kim, Moskowitz,
Dhingra & Evans (2000), and Temponi, Yen & Tiao (1999) for other
fuzzy QFD studies.
In order to address the second type of uncertainty caused
by the variability of user inputs, several studies have considered
group decision-making approaches to uncertain QFD. For example,
Ho et al. (1999) proposed an integrated group decision-making ap-
proach to consolidate individual preferences into a group consensus
by using a point-scoring scale in QFD. Karsak (2004) employed the
fuzzy Delphi method to determine the relative importance weights
of CRs based on a consensus measure and Wasserman (1993)’s pro-
cedure to normalize the fuzzy relationships between CRs and DRs.
The priority weights of DRs were then defined as the sum of the
products of the relative importance weights and the normalized
fuzzy relationships. Kwong et al. (2011) proposed a two-stage fuzzy
group decision-making approach to uncertain QFD, in which the FWA
method and FEV operator were first used to derive a ranking or-
der of DRs under each customer’s judgment, and then all customers’
preferences on the ranking of DRs were synthesized via 0-1 integer
programming to achieve a consensus ranking. Wang (2012) proposed
a different fuzzy group decision-making procedure for prioritizing
DRs under uncertainty based on FWA method and defuzzification
method, by aggregating the technical importance ratings of DRs under
each combination of each customer and each design team member.
Chen, Ko & Tseng (2013) proposed a systematic procedure for con-
structing the HOQ from a perspective of group decision-making, in
which a modified fuzzy clustering approach was proposed to find the
consensus of the design team.
In summary, all existing studies focusing on the first type of uncer-
tainty have quantified the subjective judgments in terms of fuzzy sets
and performed calculations with the associated fuzzy membership
functions of linguistic labels based on fuzzy extension principle. Con-
sequently, they have the problems of quantification, information loss,
and an obstacle to use in practice, as introduced in Section 1. More-
over, the limited studies focusing on the second type of uncertainty,
on one hand, also have the problems existed in those focusing on the
first type of uncertainty. On the other hand, they have not considered
the concept of fuzzy majority. As a result, design team’s preference
is missed in these studies. In order to overcome the drawbacks of
existing approaches, it is necessary and important to develop a more
suitable methodology for prioritizing DRs in uncertain QFD with the
consideration of the two types of uncertainties, simultaneously.
2.2. A simple and widely used framework
A full HOQ can consist of as many as 18 elements or concepts (Chan
& Wu, 2002a; Chan & Wu, 2005): customers, CRs (WHATs), correlation
matrix of WHATs, relative importance ratings of WHATs, competitors,
customer competitor assessments, goals for WHATs, sales-point, fi-
nal importance ratings of WHATs, DRs (HOWs), correlation matrix
of HOWs, relationship matrix of WHATs versus HOWs, improving
directions of HOWs, technical competitor assessment, goals for
HOWs, probability factors, importance ratings of HOWs. Due to the
high difficulty of implementing a HOQ in practice (Cristiano, Liker &
WhiteIII, 2001), it is both difficult and unnecessary to include all the
18 HOQ elements described above (Chan & Wu, 2005). In fact, dif-
ferent users build different HOQ models involving different elements
from the above list. The most simply but widely used HOQ model
contains only the customers, the CRs (WHATs) and their relative im-
portance, DRS (HOWs) and their relationships with the WHATs, and
the importance weights of the HOWs (e.g., Chen et al., 2006; Kwong et
al., 2011). According to the literature review in Section 2.1, all exist-
ing studies have the problems of quantifying qualitative concepts in
terms of fuzzy sets. Therefore, the most simply but widely used HOQ
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
[m5G;October 10, 2014;12:11]
3
model will be used in this paper to facilitate practical applications
and to clearly show the main differences between our proposed ap-
proach and existing approaches. This, however, does not imply other
factors’ unimportance in the HOQ process and future effort should be
made to handle them properly. As an extension, later in Section 6, we
will elaborate more on these omitted parts and also indicate possible
ways to include them into an even more complete HOQ model.
The first step in our QFD context is to identify the customers of
a product or service to be surveyed in a targeted market. The pro-
ducing firm should know who are the real customers for the product
concerned. There are generally three types of customers (Chan & Wu,
2002a; Chan & Wu, 2005): internal customers such as shareholders,
managers and employees; intermediate customers such as wholesale
people and retailers; and ultimate customers such as recipients of ser-
vice, purchasers, institutional purchasers. Usually the main focus is on
the ultimate customers who could be identified through previous in-
formation and marketing research. Formally, let C = {C1 , C2 , . . . , CK }
be a set of customers, where |C | = K.
Secondly, it is important for the firm to understand what cus-
tomers need for a product, otherwise you cannot know how to satisfy
your customers and thus how to keep your business successful. The
CRs (WHATs) are usually expressed in customers’ words that can be
collected by various methods by the firm’s marketing department, or
by an outside marketing or information gathering agency. Available
methods to collect CRs include focus group, individual interviews,
listening and watching, and using existing information (Chan & Wu,
2005). It is suitable and economical to gather CRs (WHATs) through
focus group and individual interviews of the K customers. It is of great
help to group related CRs into a category in analyzing the CRs. Affin-
ity diagram and cluster analysis may be used to arrange random data
into natural and logical groups so as to organize CRs. Formally, let
CR = {CR1 , CR2 , . . . , CRM } be a set of WHATs, which could be classi-
fied into some meaningful categories according to practical situations.
Thirdly, the WHATs are usually of different degrees of importance
and it is a common practice for the firm to focus more on the important
WHATs. Therefore, the importance of CRs has then to be determined to
continue the QFD process. This stage is essential because consequent
technical DRs (HOWs) are determined based on the importance of
CRs. It is usually difficult for the design team to directly assign priority
measures to a list of customer voice (CRs). Therefore, the K customers
are asked to provide their judgments for the importance of CRs. In this
step, focus groups and individual interviews are usually not suitable
for collecting quantitative information about the importance ratings
of CRs due to their high costs (Chan & Wu, 2002a). One appropriate
way of obtaining customers’ perceptions is by a survey study involving
the K customers according to their direct experiences (Temponi et al.,
1999). In order to help the K customers reveal
 their perceptions
 about
the product, a linguistic variable L1 = L11 , . . . , L1G with L11 < · · · <
1
L1G is used, e.g., {Very low, Low, Moderate, High, Very high} (Chan &
1
Wu, 2005). Formally, the importance judgment for customer need m
provided by customer k can be expressed by CRImk ∈ L1 , as shown in
Table 1 (indexed by “Importance ratings”).
Remark 1. In general, a sufficient number of customers should be
surveyed to provide a statistical significance (Chan & Wu, 2002a).
However, only a small number of customers are asked to provide
their judgments for the importance of WHATs in the literature, e.g.,
K = 5 (Chan & Wu, 2005; Liu & Wu, 2008) and K = 10 (Chen et al.,
2006). The main reason may come from the fact that the K customers
are the ones in the focus group to identify the CRs. A focus group usu-
ally consists of 8–12 randomly selected customers, who reveal their
perceptions about the product according to their direct experiences.
Consequently, the K customers may not provide a statistical signifi-
cance, but may be viewed as “experts”. In fact, Kwong et al. (2011)
and Wang (2012) have treated the HOQ as a group decision-making
problem due to the small number of customers in the survey.
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

4 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15

Table 1
A simple but widely used QFD framework.

Customers DRs

C1 C2 ... CK CRs DR1 DR2 ... DRN

Importance ratings CRI11 CRI12 ... CRI1K CR1 R11 R12 ... R1N

CRI21 CRI22 ... CRI2K CR2 R11 R12 ... R1N

. . .. . . . . . ..
. . . . . . . . .
. . . . . . .
CRIM1 CRIM2 ... CRIMK CRM R11 R12 ... R1N
Relationships
between CRs and DRs

Fig. 2. Flowchart of the proposed method.

Fourthly, the firm’s technicians or product development team
should develop a set of DRs (HOWs) to capture the WHATs in measur-
able and operable technical terms. The DRs could be generated from
current product standards or selected by ensuring through cause-
effect analysis that the DRs are the first-order causes for the CRs. For-
mally, a set of DRs collected from the design team can be expressed
as DR = {DR1 , . . . , DRN }, where N = |DR|. If necessary, these HOWs
could also be organized into some manageable categories by means
of the affinity diagram or cluster analysis.
Finally, translations from CRs to technical DRs are carried out by
a QFD design team. This is an important work in HOQ/QFD which is
performed carefully and collectively by technicians. The relationship
between a DR and a CR is usually determined by analyzing to what
extent the DR could technically relate to and influence  the CR by
making use of a linguistic variable L2 = L21 , . . . , L2G with L21 < · · · <
2
L2G , where G2 is the cardinality of L2 . The relationships can be positive
2
or negative, strong or weak. All these relationships form a matrix
with the CRs as rows and the DRs as columns. Formally, the linguistic
judgment for the relationship between customer need m and design
requirement n given by the design team is denoted by Rmn ∈ L2 , as
shown in Table 1 (indexed by “Relationships between CRs and DRs”).
With the linguistic evaluation data collected, we then aim at devel-
oping a suitable computing method that allows for deriving a priority
vector of the DRs, so as to support the prioritization decision of un-
certain QFD. A novel group decision-making approach consisting of
two stages will be proposed, the flowchart of which is depicted in
Fig. 2. The first stage is to determine the fuzzy preference relations of
different DRs with respect to each customer based on the order-based
semantics of linguistic information. The second stage is to determine
the prioritization of DRs by synthesizing all customers’ fuzzy pref-
erence relations into an overall representative one by means of the
concept of fuzzy majority.
respectively. We assume a matrix of fuzzy preference relations is
derived from the linguistic judgments so as to eliminate the burden of
quantifying qualitative concepts in uncertain QFD, which essentially
underlies the calculating basis of order-based semantics of linguistic
labels. Note that the fuzzy preference relations are employed to reflect
an omnipresent fact that the preferences may be not clear-cut so that
conventional non fuzzy preference relations may be not adequate
(Kacprzyk et al., 2008).
Looking at Table 1, we have a set of M customer needs and a set
of K customers. The customers are asked to provide their importance
judgements for different CRs, respectively. The importance judgment
for customer need m provided by customer k is denoted by CRImk ∈ L1 .
Then, a customer Ck ’s individual fuzzy preference relation in CR × CR
assigns a value in the unit interval [0, 1] for the preference of one
customer need over another such that
μDIk : (CRm , CRl ) ∈ CR × CR −→ DIkml ∈ [0, 1], (1)
where m, l = 1, 2, . . . , M, k = 1, 2, . . . , K. The value DIkml
reflects the
degree of fuzzy preference relation of CRm over CRl under Ck ’s sub-
jective judgments, defined as
⎧
⎨1, if CRImk > CRIlk
DIkml = 0.5, if CRImk = CRIlk . (2)
⎩
0, if CRImk < CRIlk
Such a value function is in fact based on the order-based semantics of
linguistic information, and consequently we will obtain K matrices of
fuzzy preference relations in CR × CR.
Then, for customer k a weighting vector, Wk , associated with the
M CRs can be induced from the matrix DIk as
 
Wk = W1k , W2k , . . . , WM
k
(3)
k is the normalized importance weight of customer need m
where Wm
under Ck ’s judgments and calculated by the following function

1
M
3. Deriving fuzzy preference relations from linguistic judgments k DIkml
Wm =
M , m = 1, 2, . . . , M, k = 1, 2, . . . , K. (4)
M
l=1 DIk
n=1 nl
In our uncertain QFD context, the linguistic information is used to
represent the subjective judgments collected from both the surveyed Remark 2. In Eq. (4), the normalized principal eigenvector is used to
customers in a targeted market and the QFD design team in a firm, derive a weighting vector of the CRs under each customer. In fact, the

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS
H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
multiplicative preference relations play a basic role in the AHP (Saaty,
1986). If DIk is a matrix of multiplicative preference relations, DIknl
should be a preference ratio such that DIknl = 1k , ∀n, l = 1, . . . , M, a
DIln
 
k k k
consistent weighting vector Wk = W1 , W2 , . . . , WM associated with
the matrix DIk should verify the following condition:
Wnk
DIknl = k
, ∀n, l = 1, . . . , M.
Wl
One commonly used method to derive the weighting vector from
a multiplicative reciprocal preference matrix is the normalized prin-
cipal eigenvector.
When DIk is a matrix of fuzzy reciprocal preference relations such
that DIknl + DIkln = 1, ∀n, l = 1, . . . , M, a consistent weighting vector
Wk associated with DIk should verify the following additive transitive
condition:
Wnk − Wlk = DIknl − 0.5, ∀n, l = 1, . . . , M.
In practice, it is quite difficult to obtain a perfectly consistent matrix
of fuzzy preference relations. Fortunately enough, Xu, Da & Liu (2009)
have proved that even if a matrix is not additive transitive, i.e., not
perfectly consistent, we can also derive the weighting vector by means
of the normalized principal eigenvector.
In addition, we also have a set of N design requirements DR =
{DR1 , DR2 , . . . , DRN }. The QFD design team is asked to provide its
subjective judgments for the relationships between different CRs and
different DRs. Similar to the process of deriving individual matrices of
fuzzy preference relations in CR × CR, we can also derive a matrix of
individual fuzzy preference relations in DR × DR under a customer
need m as
[m5G;October 10, 2014;12:11]
5
• when Dknl ∈ (0.5, 1), it indicates that DRn is slightly preferred to
DRl ;
• when Dknl = 0.5, it indicates that there is no preference (i.e. indif-
ference) between DRn and DRl ;
• when Dknl ∈ (0, 0.5), it indicates that DRl is slightly preferred to
DRn ;
• and when Dknl = 0, it indicates that DRl is absolutely preferred to
DRn .
4. Group aggregation of individual fuzzy preference relations
based on fuzzy majority
The K customers in QFD can be viewed as multiple information
sources, each of which provides a different perspective. With the
K matrices of fuzzy preference relations in DR × DR obtained, our
uncertain QFD transforms to a group decision-making problem, which
needs to synthesize the K matrices into an overall representative one.
In this section, we shall propose our prioritization solution based
on the concept of fuzzy majority. Before doing so, we first briefly
recall some basic concepts of fuzzy majority based on the exponential
generating functions (Liu, 2006).
4.1. Fuzzy majority
One of the main problems in group decision-making is to define
a decision strategy which takes into account the individual opinions
of the experts and to synthesize them into an overall representative
opinion. The reduction of the experts’ opinions into an overall repre-
sentative one is usually performed through an aggregation process.
The concept of majority is a basic element underlying group decision-
making: what is often needed is an overall opinion which synthesizes
  the opinions of the majority of the experts. The term “majority” in-
DIIm = DIIm
nl N×N dicates that a solution is accepted by most of its members, since in
where DIIm practice it is quite difficult for the solution to be accepted by all. In
nl denotes the fuzzy preference relation of DRn over DRl
with respect to customer need CRm , and calculated by the following this case the goal is to obtain a value which can be considered as the
function opinion of a majority, that is, a value that is similar for any large group
⎧ of experts, what we can call the majority opinion.
⎨1, if Rmn > Rml The concept of fuzzy majority is used to make the strict concept
DIIm
nl = 0.5, if Rmn = Rml . (5)
⎩ of majority more vague so as to make it closer to its real human
0, if Rmn < Rml perception (Kacprzyk et al., 2008). A natural manifestation of such
 
a “soft” majority is the so-called linguistic quantifiers as, e.g., most,
With the weighting vector Wk = W1k , W2k , . . . , WM
k of the M cus-
at least half, as many as possible, etc., so as to express the quantity
tomer requirements, we are able to derive a matrix of individual fuzzy of the experts which one wants to take into account. Zadeh (1983)
preference relations in DR × DR with respect to customer k (Ck ) as suggested a formal representation of these linguistic quantifiers using
follows: fuzzy sets (Zadeh, 1965), i.e., any relative quantifier can be expressed
DR1 DR2 . . . DRN as a fuzzy subset Q of the unit interval [0, 1]. In this representation
DR1 0.5 Dk12 . . . Dk1N for any proportion x ∈ [0, 1], Q (x) indicates the degree to which x
Dk = DR2 D21 0.5 . . . D2N
k k satisfies the concept conveyed by the term Q. Yager (1996) further
(6)
.. .. .. . . . defined a Regular Increasing Monotone (RIM) quantifier to represent
. . . . .. the linguistic quantifier. A fuzzy subset Q of the universe domain [0, 1]
DRN DkN1 DkN2 . . . 0.5 is called a RIM quantifier if
where k = 1, . . . , K and Q (0) = 0, Q (1) = 1, and Q (x) ≥ Q (y) for x ≥ y.

M
Given a RIM quantifier Q, we can associate with this quantifier a
Dknl = k
Wm · DIIm
nl , ∀n, l = 1, . . . , N. (7) degree of orness (Yager, 1988), defined as
m=1  1
It is easy to obtain that the matrix Dk of fuzzy preference relations orness(Q ) = Q (x)dx
satisfies the fuzzy reciprocal property such that Dknl + Dkln = 1, ∀n, l = 0

1, . . . , N, k = 1, . . . , K. Obviously, there will be K matrices of fuzzy = α (8)

preference relations in DR × DR. Such a formulation is motivated by which is equal to the area under a RIM quantifier. The semantics as-
Kwong et al. (2011) and Wang (2012) in order to take into account the sociated with the parameter α depends upon the application at hand.
group behaviors of QFD. The matrix Dk = Dknl of fuzzy preference When using the quantifier Q to approach logical connectives between
N×N
relations in DR × DR has the following properties: the and and or, the value α can be associated with a measure of the de-
gree of orness associated with an aggregation. When the quantifier Q
• when Dknl = 1, it indicates that DRn is absolutely preferred to DRl , is used to implement different procedures for approaching subjective
i.e. indicates the maximum degree of preference of DRn over DRl ; attitudes in decision-making under uncertainty, α can be associated

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS
6 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
with degree of optimism of a decision-maker. Since we want to model
the group behaviors of uncertain QFD, α is viewed as the optimism
index of the uncertain QFD planning process. In this context,
• when α < 0.5 it indicates that the decision-maker has a pes-
simistic attitude toward the reliability of customers, the value
α = 0.0 means that the decision-maker is absolutely pessimistic;
• when α = 0.5, it indicates that the decision-maker has a neutral
attitude toward the reliability of customers;
• and when α > 0.5 it means that the decision-maker has an op-
timistic attitude toward the reliability of customers, the value
α = 1.0 means the decision-maker is absolutely optimistic.
Obviously, it is an important issue to define the membership func-
tion for a quantifier in group decision-making. Until now, a quan-
tifier’s membership function is often determined by intuition (e.g.,
Zadeh, 1983; Yager, 1988). The main drawbacks of such a way are that
the quantifiers defined by intuition are often incomparable and the
aggregation results are inconsistent in real applications (Liu, 2006).
Recently, Liu (2006) has proposed a way to derive the membership
function of a RIM quantifier Q from the orness value α based on the
exponential generating function of a RIM quantifier, which will used
in this paper. The readers can refer to Appendix A for more detail
about the exponential RIM quantifier and its implementation.
4.2. Aggregation of fuzzy preference relations with fuzzy majority
We now consider how to derive our prioritization solution based
on the concept of fuzzy majority, which is guided by the exponential
RIM linguistic quantifier Q. In our context, the linguistic quantifier Q
expresses the quantity of the customers which the QFD design team
wants to take into account. A linguistically quantified statement for
our QFD problem can be written as “QCs are convinced”, which may
be exemplified by
“The QFD design team is α optimistic that customers (Cs) are
convinced ()”.
Such a linguistic statement can be solved by an aggregation func-
tion. To solve the linguistic statement (), we will build our group
aggregation based on Kacprzyk et al. (2008) following the scheme:
{D1 , . . . , DK } → D → Prioritization of DRs,
i.e., from the individual fuzzy preference relations we determine first
a social fuzzy preference relation, D, which is similar in spirit to its
individual counterpart but concerns the whole group of individuals,
and then find a solution from such a social fuzzy preference relation.
Firstly, we assume a most straightforward alternative that the
social fuzzy preference relation D = [Dnl ]N×N is given by
1 k
K
Dnl = Dnl .
K
k=1
where n, l = 1, . . . , N.
Secondly, a solution concept of much intuitive appeal is here the
consensus winner which will be extended under a social fuzzy prefer-
ence relation and a fuzzy majority. We start with
1
N
Hn = Hnl , n = 1, . . . , N
N−1
l=1,l=n
where
 
1, if Dnl > 0.5;
Hnl =
0, otherwise.
The value Hnl expresses whether design requirement DRn defeats (in
the whole group’s opinion) design requirement DRl or not. The value
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
[m5G;October 10, 2014;12:11]
Hn is the mean degree to which design requirement DRn is preferred
to all the other DRs.
Thirdly, by taking into account the exponential RIM linguistic
quantifier Q, we calculate
VQn = μQ (Hn ), (13)
which is the extent, from 0 to 1, to which design requirement DRn is
preferred by the whole customers, over Q other DRs. We then define
the fuzzy Q-consensus winner as
VQ1 VQN
VQ = + ··· + , (14)
DR1 DRN
i.e., as a fuzzy set of DRs that are preferred, by all customers, over Q
other DRs. Such a fuzzy set is used to prioritize the DRs. The greater the
value of VQn , the higher priority of design requirement DRn . It should
be noted here that a threshold value B may be specified by the design
team to prioritize the design requirements (Vanegas & Labib, 2001).
4.3. Consensus degrees in uncertain QFD
Now we consider how to utilize fuzzy majority in terms of the
exponential RIM linguistic quantifier Q to define a consensus degree
of our uncertain QFD problem. Such a consensus degree is borrowed
from Kacprzyk et al. (2008) and meant to overcome some rigidity of
the conventional concept of consensus, in which full consensus occurs
only when “all the customers agree on the solution”.
Our starting point to the consensus degree is again the K matrices
Dk of fuzzy preference relations in DR × DR with respect to the
K customers, as obtained in Section 3. We start with the degree of
agreement between two customers Ck1 and Ck2 as to their preferences
between DRn and DRl , defined as follows.
k k
Definition 1. Let Ck1 , Ck2 be two customers, Dnl1 , Dnl2 be the fuzzy
preference relation of DRn over DRl under Ck1 and Ck2 , respectively.
The degree of agreement between Ck1 and Ck2 with respect to the
fuzzy preference relation of DRn over DRl is defined as
   
 
Agreenl Ck1 , Ck2 = 1 − Dknl1 − Dknl2  (15)
where here and later on in this section, if not otherwise specified, k1 =
1, . . . , K − 1; k2 = k1 + 1, . . . , K; n = 1, . . . , N − 1; l = n + 1, . . . , N.
(9)
The agreement measure between Ck1 and Ck2 is based on the de-
viation degree and has
 a definite  physical implication. Obviously, the
 k k 
smaller the value of Dnl1 − Dnl2 , the greater the value Agreenl (Ck1 , Ck2 ),
and the more similar these two customers Ck1 , Ck2 with respect to the
fuzzy preference relation of DRn over DRl . Since D·nl ∈ [0, 1], it is easy
to get Agreenl (Ck1 , Ck2 ) ∈ [0, 1]. With respect to the fuzzy preference
relation of DRn over DRl , it is easily found that
(10) • if Agreenl (Ck1 , Ck2 ) = 1, then the two customers Ck1 , Ck2 have no
difference (full agreement);
• if Agreenl (Ck1 , Ck2 ) = 0, then the two customers Ck1 , Ck2 are abso-
lutely different (full disagreement).
 
Secondly, we aggregate these degrees Agreenl Ck1 , Ck2 to obtain a
degree of agreement of each pair of customers as to their preferences
between all the pairs of DRs, defined as follows.
(11)
Definition 2. The degree of agreement between customers Ck1 and
Ck2 as to their preferences between all the pairs of DRs is defined as

Agree Ck1 , Ck2
(12) 

 2
N=1 N
= [Agreenl (Ck1 , Ck2 )] .
2
(16)
N(N − 1)
n=1 l=n+1
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15 7

Table 2
A fried Chinese vegetable: Five customers’ perceptions on the relative importance of the 10 WHATs.

C WHATs (CRs)

CR1 CR2 CR3 CR4 CR5 CR6 CR7 CR8 CR9 CR10

C1 High Moderate Very high High Low Very high Moderate High Low Very low
C2 Very high Very high Moderate Low Very low High Low Moderate Moderate Low
C3 High Moderate Very high Moderate Low High Very high Moderate Very low Moderate
C4 Moderate High Moderate Very high Moderate Moderate High High Low Very low
C5 Very high Moderate Low Low High Very high Moderate High Moderate Low

 
It is obvious that Agree Ck1 , Ck2 is also based on the deviation of edible oil used” and “duration of vegetable storage”) and then to
degree. The greater the value Agree(Ck1 , Ck2 ), the more similar these find the important ones through technical analyses.
two customers Ck1 , Ck2 . At first, the restaurant must know who are the customers for its
Then, the degree of agreement of all the pairs of customers as to fried Chinese vegetable. This information could be obtained by ob-
their preferences between pairs of DRs is defined as serving who order the fried Chinese vegetable or through market
 survey. Five of the restaurants’ customers are selected to help con-


K−1
K
  2
 2
Agree =  Agree Ck1 , Ck2
K (K − 1)
k1 =1 k2 =k1 +1
Finally, the degree of agreement of all the pairs of customers as to
their preferences between Q pairs of DRs is defined as
 
AgreeQ = μQ Agree .
5. Illustrative application case studies
In this section, we shall apply our proposed approach to prioritize
the basic HOWs (DRs) of two examples borrowed from the literature.
Our main attempt here is to present a generalized model based on
past studies that can then be modified or extended for use in a specific
situation or firm. Moreover, the examples provide illustrations of the
steps required to apply the methodology in practice; they can be
summarized as follows.
• Step 1: Determine customers and customer needs (CRs/WHATs).
The customers are then asked to linguistically rate the importance
of the WHATs.
• Step 2: Determine HOWs and ask the design team to linguistically
judge the relationships between HOWs and WHATs.
• Step 3: With the importance ratings and relationships between
HOWs and WHATs, as shown in Table 1, prioritize the DRs (HOWs)
based on the proposed approach.
– Step 3.1: Calculate the normalized importance weights of dif-
ferent WHATs under each customer via Eq. (4).
– Step 3.2: Derive a matrix of individual fuzzy preference re-
lations of different HOWs with respect to each customer via
Eq. (6).
– Step 3.3: Induce the prioritization of the HOWs via Eqs. (10)–
(14).
– Step 3.4: Calculate the consensus degree via Eqs. (15)–(18).
5.1. Example 1: A Chinese restaurant
The example of discussing the basic HOWS (DRs) of a fried Chinese
vegetable (Chan & Wu, 2005) in a Chinese restaurant is first revised
here to illustrate the concepts and computations in our proposed HOQ
model in details. A Chinese restaurant, called restaurant, wishes to
make an improvement on a fried Chinese vegetable it cooks and sells
everyday in response to the competition of other Chinese restaurants
in the same district. HOQ technique can help make the appropriate
decision resulting in better improvement. The basic idea is (1) to un-
derstand what are customer needs for a fried Chinese vegetable (such
as “not greasy” and “fresh”) and then to identify the important ones
through customer surveys, and (2) to associate the customer needs
with appropriate technical measures or solutions (such as “amount
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
duct the HOQ analysis. By one or more focus groups, these five cus-
(17)
tomers identify 10 needs (WHATs) for the fried Chinese vegetable.
They are: “CR1 : Appetizing”, “CR2 : Jade color”, “CR3 : Fresh”, “CR4 :
Hot (High temperature)”, “CR5 : Moderately spicy”, “CR6 : Not greasy”,
“CR7 : Not overcooked”, “CR8 : Tasteful arrangement”, “CR9 : Not salty”,
and “CR10 : Sweet smell”. The 10 WHATs can hardly be of same impor-
(18) tance to the customers. Consequently, the five customers are asked to
reveal their perceptions on the relative importance of the 10 WHATs
using the following linguistic variable
L1 = {Very low, Low, Moderate, High, Very high}. (19)
Table 2 shows their assessments where, for example, customer 1 (C1 )
rates the relative importance of customer need 1 (CR1 ) as “High”.
After careful considerations, restaurant’s cooks (design team) pro-
pose nine technical measures (HOWs) that relate to and can help
realize the 10 WHATs. These HOWs are “DR1 : amount of soy sauce
used”, “DR2 : amount of salt used”, “DR3 : amount of pepper used”,
“DR4 : amount of vegetable used”, “DR5 : amount of edible oil used”,
“DR6 : duration of vegetable storage”, “DR7 : duration between fin-
ished and served”, “DR8 : duration of cooking”, and “DR9 : duration of
high temperature used in cooking”. Then the cooks begin to establish
the relationships between the HOWs and the WHATs, or to examine
to what extent each HOW is related to and can technically measure
each WHAT. Each HOW’s relationships with the WHATs are always
examined once the HOW is considered. The relationships between
the HOWs and the WHATs are determined by technical analysis and
empirical judgment with the following linguistic variable
L2 = {Very weak, Weak, Moderate, Strong, Very strong}. (20)
The full matrix of these relationships are shown in Table 3 where, for
example, the cooks consider the relationship between DR1 and CR1
as “Very strong”.
With the linguistic judgments provided by the five customers and
the cooks (QFD design team), we are able to derive the matrix of
fuzzy preference relations for each customer, respectively. Firstly,
we calculate a matrix DIk of fuzzy preference relations of different
WHATs (CRs) from customer k ( Ck )’s subjective
 judgments, according
to Eq. (2). Then, a weighting vector Wk = W1k , . . . , W10
k associated
with the 10 WHATs can then be derived from the matrix DIk with re-
spect to customer k, according to Eq. (4). Table 4 shows the weighting
vectors of the 10 WHATs with respect to the five customers.
Secondly, we can obtain a matrix DIIm of fuzzy preference relations
of the nine HOW (DRs) from the cooks’ subjective judgements for the
relationships between CRs and DRs with respect to customer need m.
Taking the weighting vector Wk into account, we can obtain a matrix
Dk of the fuzzy preference relations of DRs with respect to customer
k according to Eq. (7). Taking customer 1 (C1 ) as an example, the
derived matrix of fuzzy preference relations of the nine DRs is derived
as shown in Table 5.
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

8 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15

Table 3
A fried Chinese vegetable: Cooks’ linguistic judgments on the relationship matrix of HOWs versus WHATs.

DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9

CR1 Very strong Strong Very weak Very weak Moderate Weak Very weak Strong Moderate
CR2 Strong Very strong Moderate Moderate Moderate Moderate Moderate Moderate Strong
CR3 Strong Strong Weak Very weak Moderate Very weak Weak Strong Moderate
CR4 Strong Strong Very weak Very weak Moderate Very weak Weak Moderate Strong
CR5 Weak Weak Very strong Very strong Strong Weak Weak Moderate Moderate
CR6 Moderate Moderate Weak Weak Very strong Weak Weak Moderate Moderate
CR7 Weak Weak Very weak Very weak Strong Very weak Very weak Weak Weak
CR8 Moderate Moderate Weak Weak Strong Very strong Strong Weak Moderate
CR9 Very weak Very weak Very weak Very weak Moderate Strong Very strong Very weak Very weak
CR10 Very weak Very weak Very weak Weak Weak Moderate Moderate Strong Strong

Table 4
A fried Chinese vegetable: Importance weights of the 10 CRs under each customer.

CR1 CR2 CR3 CR4 CR5 CR6 CR7 CR8 CR9 CR10

C1 0.13 0.08 0.18 0.13 0.04 0.18 0.08 0.13 0.04 0.01
C2 0.18 0.18 0.11 0.05 0.01 0.15 0.05 0.11 0.11 0.05
C3 0.14 0.08 0.18 0.08 0.03 0.14 0.18 0.08 0.01 0.08
C4 0.08 0.15 0.08 0.19 0.08 0.08 0.15 0.15 0.03 0.01
C5 0.18 0.09 0.03 0.03 0.14 0.18 0.09 0.14 0.09 0.03

Table 5
A fried Chinese vegetable: Fuzzy preference relation of different DRs under customer C1 (D1 ).

DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9

DR1 0.5 0.525 0.935 0.93 0.52 0.8 0.8 0.71 0.63
DR2 0.475 0.5 0.935 0.93 0.52 0.8 0.8 0.645 0.67
DR3 0.065 0.065 0.5 0.585 0.08 0.455 0.365 0.165 0.06
DR4 0.07 0.07 0.415 0.5 0.085 0.365 0.275 0.165 0.06
DR5 0.48 0.48 0.92 0.915 0.5 0.78 0.845 0.575 0.625
DR6 0.2 0.2 0.545 0.635 0.22 0.5 0.455 0.21 0.17
DR7 0.2 0.2 0.635 0.725 0.155 0.545 0.5 0.21 0.17
DR8 0.29 0.355 0.835 0.835 0.425 0.79 0.79 0.5 0.485
DR9 0.37 0.33 0.94 0.94 0.375 0.83 0.83 0.515 0.5

Thirdly, we perform the group aggregation of the K individual ma-
trices (Dk , k = 1, . . . , 5) into an overall representative one by means of
the solution introduced in Section 4.2. We first consider the aggrega-
tion without using the concept of fuzzy majority, the priority values
are shown as
H1 H2 H3 H4 H5 H6 H7 H8 H9
(21)
0.875 0.75 0.125 0.0 1.0 0.25 0.375 0.5 0.625
which indicates the following prioritization of the nine HOWs: DR5 
DR1  DR2  DR9  DR8  DR7  DR6  DR3  DR4 , i.e., the priority
value of DR5 is 1.0, which indicates that DR5 is the most important
design requirement; whereas DR4 is the most unimportant one with
a priority value 0.
Now, we consider incorporating the exponential RIM quantifier
to model the concept of fuzzy majority. Consequently we can obtain
different prioritization information of the nine HOWs under different
α values, as shown in Table 6, which has the following interesting
properties
• When α = 0, the priority values of DRs are always 0 except the one
of DR5 . The main reason is that in this case the cooks are absolutely
pessimistic and only the priority value of DR5 is equal to 1. In this
case, only one important HOW can be selected from DR, i.e., the
cooks only need to pay attention to DR5 .
• When α = 1, DR4 is the most unimportant design requirement,
whereas the remaining HOWs (DRs) are all the most important
ones with a priority value 1. The main reason is that in this case
the cooks are in fact absolutely optimistic and the priority values
of the HOWs are greater than 0 except the one of DR4 .
• When α = 0.5, the final priority information with fuzzy majority
is the same as the one without using fuzzy majority. The main
reason is that in this case the cooks in fact have a neutral attitude
and the quantifier is defined as Q (x) = x, which will be provided
as a default recommendation to the cooks.
• When α ∈ (0, 0.5) or α ∈ (0.5, 1), the priority ranking of the nine
HOWs is DR5  DR1  DR2  DR9  DR8  DR7  DR6  DR3 
DR4 . Since the exponential RIM quantifier Q is monotonically non-
decreasing with α , the priority information with respect to each
design requirement is monotonically non-decreasing with α .
In addition, it is observed from Table 6 that different priority val-
ues can be obtained with respect to different α values, which reflect
the cooks’ attitude toward the reliability of the information provided
by the five customers. Such a change from the cooks’ attitude may
play the following two roles in QFD. Firstly, our research context is
the first step of the four-stage QFD, the priority information will be
transferred to the second stage. Therefore, the priority information
under different α values will inevitably influence the second stage.
Secondly, in practice it is not so easy for the cooks to pay attention
to all the important HOWs. Under their attitude α , they can choose
the most important DRs by specifying a threshold value B (Vanegas
& Labib, 2001). In this case, the number of important DRs may be
different with respect to α . Setting B = 0.5 as an example, the pri-
ority rankings of important DRs under different α values are shown
in Table 7, which shows that the number of important DRs (HOWs)
increases with the cooks’ attitudinal character α .
Finally, given an α value a consensus degree can be obtained ac-
cording to Eqs. (15)–(18). To clearly show the relationship between
the consensus degree and an α value, the consensus degrees with
respect to different α values are depicted in Fig. 3(a), which shows
that the greater value the α , the higher consensus degree, i.e., the
consensus degree is monotonically non-decreasing with α .

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15 9

Table 6
A fried Chinese vegetable: Prioritization information of the nine DRs under different α values.

RIM quantifiers with α : VQα

DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9

α = 0.0 0.0 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0
α = 0.1 0.2866 0.0821 0.0001 0.0 1.0 0.0005 0.0019 0.0067 0.0235
α = 0.2 0.545 0.2953 0.0068 0.0 1.0 0.0192 0.0419 0.0831 0.1583
α = 0.3 0.695 0.4765 0.0294 0.0 1.0 0.0706 0.128 0.2082 0.3201
α = 0.4 0.7986 0.626 0.0686 0.0 1.0 0.1487 0.2421 0.3509 0.4779
α = 0.5 0.875 0.75 0.125 0.0 1.0 0.25 0.375 0.5 0.625
α = 0.6 0.9314 0.8513 0.2014 0.0 1.0 0.374 0.5221 0.6491 0.7579
α = 0.7 0.9706 0.9294 0.305 0.0 1.0 0.5235 0.6799 0.7918 0.872
α = 0.8 0.9932 0.9808 0.455 0.0 1.0 0.7047 0.8417 0.9169 0.9581
α = 0.9 0.9999 0.9995 0.7134 0.0 1.0 0.9179 0.9765 0.9933 0.9981
α = 1.0 1.0 1.0 1.0 0.0 1.0 1.0 1.0 1.0 1.0

Table 7 objective means of assuring that requirements have been met. The
A fried Chinese vegetable: Priority rankings under different α values
relationship measure between CRs and DRs are assessed by the QFD
with benchmark B = 0.5.
design team with the following linguistic variable
RIM quantifiers with α : VQα
L2 = {Very weak (VW), Weak (W), Moderate (M), Strong (S)}.
Priority rankings of important DRs

α = 0.0 DR5 Table 8 (indexed by “Relationships between CRs and DRs”) shows the
α = 0.1 DR5 linguistic relationship matrix between the eight CRs and the 10 DRs
α = 0.2 DR5  DR1 assessed by the QFD design team where, for example, the design team
α = 0.3 DR5  DR1
considers the relationship between DR1 and CR1 as “M (Moderate)”.
α = 0.4 DR5  DR1  DR2
α = 0.5 DR5  DR1  DR2  DR9  DR8
The application procedure of our proposed approach to this FMS
α = 0.6 DR5  DR1  DR2  DR9  DR8  DR7 example is the same as the one to the restaurant example, therefore
α = 0.7 DR5  DR1  DR2  DR9  DR8  DR7  DR6 we here only give the final results of this FMS example. Table 9 shows
α = 0.8 DR5  DR1  DR2  DR9  DR8  DR7  DR6 different priority information of the 10 HOWS (DRs) under different
α = 0.9 DR5  DR1  DR2  DR9  DR8  DR7  DR6  DR3
α = 1.0 DR5  DR1  DR2  DR9  DR8  DR7  DR6  DR3
α values, which has similar properties with the restaurant example.
As a default recommendation to the design team, priority ranking of
the 10 DRs is DR3  DR7  DR1  DR2  DR4  DR8  DR9  DR10 
DR6  DR5 .
5.2. Example 2: A flexible manufacturing system Finally, we consider the consensus degree in this FMS example. The
consensus degrees with respect to different α values are depicted in
The example discussing the basic HOWS (DRs) of a flexible man- Fig. 3(b), which has similar properties with the one (Fig. 3(a)) of the
ufacturing system (FMS; Khoo & Ho, 1996; Chen et al., 2006; Kwong restaurant example. It is also observed from Fig. 3 that the line of
et al., 2011; Wang, 2012) is also applied here to illustrate the idea this FMS example is steeper than that of the restaurant example. For
proposed. In that example, 8 major WHATs (CRs) are identified in a example, when α = 0.4 we find AgreeFMS ≈ 0.08 and AgreeRestaurant ≈
Q Q
survey market for the design of the FMS, which are “CR1 : High produc- 0.18, which indicates that customers in the restaurant example have
tion volume”, “CR2 : Short setup time”, “CR3 : Load-carrying capacity”, a higher consensus degree than these in this FMS example.
“CR4 : User-friendliness”, “CR5 : Associated functions”, “CR6 : Modular-
ity”, “CR7 : Wide tool variety”, and “CR8 : Wide product variety”. Then,
10 surveyed customers are asked to assess the eight CRs by making 5.3. Comparative analyses
use the following linguistic variable
Our uncertain QFD context has been widely investigated in the
L1 = {Very unimportant (VU), Quite unimportant (QU), literature by quantifying qualitative concepts with fuzzy sets. In this
Unimportant (U), Slightly important (SI), section, we compare our proposed approach with Chen et al. (2006),
Kwong et al. (2011), and Wang (2012), all of which have utilized the
Moderately important (MI), Important (I),
nonlinear (fractional) programming (NLP) based FWA (NLP-FWA for
Very important (VI)} short) method (Kao & Liu, 2001). Before doing so, it is necessary to
Their importance assessments are shown in Table 8 (indexed by give brief descriptions of these three approaches.
“Importance ratings”) where, for example, customer 1 rates the rela-
tive importance of CR1 as “VI (Very important)”. 5.3.1. Brief descriptions of three existing approaches
Based on the design team’s experience and expert knowledge, 10 In their seminal work of fuzzy linguistic QFD, Chen et al. (2006)
HOWs (DRs) are identified responding to the eight major CRs, which proposed a method by means of the NLP-FWA method and fuzzy
are “DR1 : Automatic gauging”, “DR2 : Tool change system”, “DR3 : Tool expected value (FEV) operator (Liu & Liu, 2002) to rank DRs in fuzzy
monitoring system”, “DR4 : Coordinate measuring machine”, “DR5 : linguistic QFD. Their approach can be summarized as follows.
Automated guided vehicle”, “DR6 : Conveyor”, “DR7 : Programmable
logic controller”, “DR8 : Storage and retrieval system”, “DR9 : Modular • Firstly, fuzzy importance weights of WHATs (CRs) and fuzzy re-
fixture”, and “DR10 : Robots”. According to the customers, the design lationships between CRs and HOWs (DRs) provided by multiple
team needs to translate the customers’ wishes into technical speci- customers and QFD team members are averaged.
fications from the DRs. These requirements must be in balance with • Secondly, the NLP-FWA method is used to determine the fuzzy
available expertise and the given time and cost frame of the project. weights of DRs.
Consequently, the design team needs to determine the priorities of • Finally, the fuzzy weights of DRs are defuzzified using the FEV
the DRs in developing a new model of an FMS and to provide an operator for prioritizing DRs.

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

10 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15

(a) (b)
Fig. 3. Consensus degrees under different α values (a) a fried Chinese vegetable; (b) an FMS.

Table 8
An FMS: Linguistic importance of the eight CRs assessed by 10 customers.

Customers DRs

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 CRs DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9 DR10

VI U VU VI MI VI VI I VI QU CR1 M W M M VW VW M W W W
QU U VI SI QU SI QU VI VU SI CR2 W M S W VW M M M M W
VI I QU VI QU VU I MI I VI CR3 S M S W VW VW S W W W
QU VI MI VU I MI VU VI U QU CR4 M M M M W W W M M W
VI VU VU U MI U SI VI VI VI CR5 S S M S W W M M M M
U QU VU I VU QU MI I MI VU CR6 W W M M W W W W W W
SI I MI VI VI VI VU QU VI I CR7 W M M W M M S M W W
VU VI VI QU I QU I QU VI QU CR8 M W M M W W S M W W

Importance ratings Relationships between CRs and DRs

Table 9
A FMS: Priority information under different α values.

RIM quantifiers with α : VQα

DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9 DR10

α = 0.0 0.0 0.0 1.0 0.0 0.0 0.0 0.0 0.0 0.0 0.0
α = 0.1 0.1085 0.0357 1.0 0.0117 0.0 0.0001 0.3294 0.0038 0.0012 0.0004
α = 0.2 0.3387 0.1952 1.0 0.1111 0.0 0.0058 0.5832 0.0617 0.0328 0.0158
α = 0.3 0.519 0.3666 1.0 0.2534 0.0 0.0257 0.7241 0.1692 0.1067 0.0602
α = 0.4 0.6621 0.5248 1.0 0.405 0.0 0.0605 0.8195 0.3004 0.2093 0.1298
α = 0.5 0.7778 0.6667 1.0 0.5556 0.0 0.1111 0.8889 0.4444 0.3333 0.2222
α = 0.6 0.8702 0.7907 1.0 0.6996 0.0 0.1805 0.9395 0.595 0.4752 0.3379
α = 0.7 0.9398 0.8933 1.0 0.8308 0.0 0.2759 0.9743 0.7466 0.6334 0.481
α = 0.8 0.9842 0.9672 1.0 0.9383 0.0 0.4168 0.9942 0.8889 0.8048 0.6613
α = 0.9 0.9996 0.9988 1.0 0.9962 0.0 0.6706 0.9999 0.9883 0.9643 0.8915
α = 1.0 1.0 1.0 1.0 1.0 0.0 1.0 1.0 1.0 1.0 1.0

Kwong et al. (2011) approach is a fuzzy group decision-making ap-
proach that integrates the NLP-FWA method with a consensus ordinal
ranking technique, summarized as follows.
• Firstly, the importance weights of HOWs (DRs) with respect to
each customer are determined using the NLP-FWA method.
• Secondly, the FEV operator is used to defuzzify the fuzzy weights
of DRs. An ordinal ranking of DRs with respect to each customer
can then be generated from the defuzzification.
• Finally, all customers’ preferences on the ranking of DRs are syn-
thesized via 0-1 integer programming to achieve a consensus
ranking.
Wang (2012) approach is a fuzzy normalization based group
decision-making approach, which can be summarized as follows.
• Firstly, the importance weights of HOWS (DRs) with respect to
each customer are determined using the NLP-FWA method.
• Secondly, the importance weights of DRs with respect to different
customers are normalized by the NLP-FWA method.
• Finally, a defuzzification method is used to prioritize the DRs.
For more details about these three existing techniques, the readers
can refer to Chen et al. (2006), Kwong et al. (2011), and Wang (2012).
All of the three approaches have been implemented by means of Java
programming language in this study. Moreover, in order to clearly

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15 11

Table 10
Prioritization results of existing approaches with Eq. (22).

Approaches Information Prioritization results

DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9

Chen et al. (2006) Value 0.541 0.5366 0.2743 0.261 0.6033 0.3623 0.3674 0.4901 0.506
Ranking DR5  DR1  DR2  DR9  DR8  DR7  DR6  DR3  DR4
Kwong et al. (2011) Value No
Ranking DR5  DR1  DR2  DR9  DR8  DR7  DR6  DR3  DR4
Wang (2012) Value 0.143 0.1419 0.0739 0.0705 0.1589 0.0969 0.0982 0.1297 0.1338
Ranking DR5  DR1  DR2  DR9  DR8  DR7  DR6  DR3  DR4

Table 11
Prioritization values of Chen et al. (2006) and Wang (2012) with Eq. (23).

Approaches Priority values

DR1 DR2 DR3 DR4 DR5 DR6 DR7 DR8 DR9

Chen et al. (2006) 0.6494 0.6455 0.3988 0.3871 0.6992 0.4748 0.475 0.5958 0.612
Wang (2012) 0.143 0.1422 0.0898 0.0874 0.1532 0.1063 0.1064 0.1317 0.1351

show the main differences of these three existing approaches, the FEV
operator is used to defuzzify a fuzzy value in Wang (2012) approach.
5.3.2. Results and comparisons
Here, the restaurant example in Section 5.1 is used as a com-
parison for the illustration purpose. As mentioned before, the above
three existing approaches have quantified the qualitative concepts
(linguistic judgments) by means of fuzzy sets. Similar with Chan and
Wu (2005), the following fuzzy linguistic variables have been used to
semantically represent the linguistic judgements.
L1 = {Very low, Low, Moderate, High, Very high}
= {(0, 1, 2), (2, 3, 4), (4, 5, 6), (6, 7, 8), (8, 9, 10)}
L 2
= {Very weak, Weak, Moderate, Strong, Very strong}.
= {(0, 0.1, 0.2), (0.2, 0.3, 0.4), (0.4, 0.5, 0.6),
(0.6, 0.7, 0.8), (0.8, 0.9, 1.0)}
where (·, ·, ·) is used to represent a triangular fuzzy number. With the
linguistic judgments in Tables 2–3 and their semantically represented
fuzzy numbers in Eq. (22), we can calculate the prioritization results
by the three existing approaches using our Java programs, as shown in
Table 10. It is found that the three existing approaches generate same
priority ranking as ours. Kwong et al. (2011) approach only generates
a priority ranking without values, it may not be suitable since the
priority values can also be transformed to the later stages.
Despite the same priority ranking generated, the above three
existing approaches make use of the associated fuzzy membership
function of every linguistic label, which is often defined subjectively
and context dependently. Such a need to define fuzzy membership
functions for the qualitative concepts may create a burden in the ap-
plication of QFD, since QFD design team may not be familiar with
the knowledge of fuzzy sets. Due to the difficulty to use QFD in prac-
tice (Cristiano et al., 2001), the three existing approaches need costly
learning processes of fuzzy sets that could reduce the usefulness of
these approaches. Instead, our approach performs computation di-
rectly with words based on the order-based semantics of linguistic
labels. Consequently, the burden of quantifying qualitative concepts
can be eliminated. In this sense, it gives a practical convenience in the
process of QFD planning.
Even if the fuzzy membership functions can be easily learned and
defined by the QFD design team, such a quantification is in fact the
process of transforming an ordinal scale into a cardinal scale that rep-
resents an “arbitrary passage”, therefore it is easy to produce different
results by choosing different scales from which to draw the ordinals
(Han et al., 2004). Similar with Liu & Wu (2008), the following fuzzy
linguistic variables can be used to quantify the qualitative concepts
in our research context.
L1 = {Very low, Low, Moderate, High, Very high}
= {(0, 0, 0.25), (0, 0.25, 0.5), (0.25, 0.5, 0.75),
(0.5, 0.75, 1), (0.75, 1, 1)}
L2 = {Very weak, Weak, Moderate, Strong, Very strong}.
= {(0.2, 0.2, 0.4), (0.2, 0.4, 0.6), (0.4, 0.6, 0.8),
(0.6, 0.8, 1.0), (0.8, 1.0, 1.0)} (23)
which is quite different from Eq. (22). With the new fuzzy mem-
bership functions defined, we can obtain the prioritization results of
the nine DRs by the three existing approaches. Interestingly, the pri-
ority rankings generated by the three approaches keep the same as
ours. Since Kwong et al. (2011) approach only generates a priority
ranking without priority values, we only list the priority values of
(22) Chen et al. (2006) and Wang (2012) in Table 11, which are quite dif-
ferent from the ones in Table 10. It is found that the priority values
change with the definition of fuzzy membership functions. More-
over, in Chen et al. (2006) results, the priority value of DR6 differs
slightly from the one of DR7 , i.e., V6 = 0.4748 and V7 = 0.475. The
same phenomenon appears in Wang (2012) results. Due to the small
difference of weights between DR6 and DR7 , the cooks may treat these
two HOWs equivalently in practice. Since the priority values may be
transformed to the later stages in QFD, quantification in terms of fuzzy
sets may sometimes be dangerous.
Due to the customer heterogeneity in QFD, it is necessary to con-
sider the group behaviors underlying the uncertain QFD problem.
In order to show this phenomenon, we slightly change the original
evaluation data in Table 2 as “CRI39 = Low, CRI59 = Moderate”, which
means that customers 3 and 5 exchange their judgments on design
requirement 9. In this case, our approach generates the prioritization
results as
H1 H2 H3 H4 H5 H6 H7 H8 H9
0.875 0.75 0.125 0.0 1.0 0.375 0.25 0.5 0.625
which indicates the following prioritization of the nine HOWs DR5 
DR1  DR2  DR9  DR8  DR6  DR7  DR3  DR4 . In this case, our
  
approach prioritizes DR6 and DR7 as DR6  DR7 , which is contrast to
the priority ranking and values generated using the original data in
Table 2. Such a small change reflects the customer heterogeneity. By
the three existing approaches implemented with our Java programs,
it is found that Kwong et al. (2011) approach generates the same pri-
ority ranking as ours, whereas Chen et al. (2006) and Wang (2012)
approaches still generate the same priority ranking as the one using
the original data in Table 2. In this sense, Kwong et al. (2011) approach

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS
12 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
can also reflect the customer heterogeneity but it does not give the
priority values. Both Chen et al. (2006) and Wang (2012) approaches
cannot reflect the customer heterogeneity. Moreover, the three exist-
= R ⊗ CorDR
ing approaches do not consider the concept of fuzzy majority, which
is an underlying concept in group decision-making. Instead, our pro-
posed approach considered the fuzzy majority by means of an expo-
nential RIM quantifier, in which an attitudinal character α would be
and R
) obtained, our QFD con-
provided by the design team, as shown in Section 5.1. Our approach
generated the same priority ranking but different priority values with
the change of α values, which reflects the cooks’ attitudinal prefer-
ence toward the customer heterogeneity. Since our context is the first
step of the four-stage QFD, the priority values under different α val-
ues will inevitably influence the second stage. Motivated by Vanegas
and Labib (2001), the cooks can choose different important DRs under
different α values by specifying a benchmark value.
6. About other elements in our QFD context
As mentioned in Section 2.2, there are other elements in a full
QFD. For the completeness of this study, here we provide some con-
siderations about them and suggest some ways to handle them in a
potentially more comprehensive HOQ model. According to the HOQ
model in Fig. 1, the other elements are classified into three groups:
correlations among WHATs and among HOWs, planning matrix, and
technical matrix. Since this paper focuses on uncertain QFD with lin-
guistic judgments, other factors involving linguistic information are
considered in this section. The elements in terms of crisp numerical
representation can be easily revised and incorporated into our QFD
 the M WHATs with respect to customer k is denoted
context according to Chan and Wu (2005) procedures, and thus will
not be discussed in this study.
 the M WHATs under each
6.1. Correlations among WHATs and among HOWs
Correlations among the HOWs are mentioned and included in
many QFD studies, but are seldom incorporated into the calculation
of the HOWs’ final ratings. Very few studies include or even mention
the correlations among the WHATs, not to say incorporate such cor-
relations into the calculation of the WHATs’ final ratings. One reason
for these is the difficulty in obtaining the relevant data since, although
they can be measured using appropriate linguistic variables, correla-
tions among the HOWs and among the WHATs require the technicians
and the customers to make a lot of pair-wise comparisons about the
degrees to which the HOWs and the WHATs are inter-related. This
would be tedious and difficult from the customers’ perspective, al-
though it is possible from the technician’s side. Among the very few
approaches found in the QFD literature to deal with the correlations,
a noticeable one is proposed by Khoo and Ho (1996) to determine its
importance rating as a liner combination of its correlations with other
WHATs weighted by the WHATs’ initial importance weights. Since we
focus on group approach to QFD based on the order-based semantics
of linguistic labels, the potential incorporation of the correlations
into our QFD context may consider linguistic aggregation. Formally,
let CRI = [CRIkm ]K×M be the transformed matrix of the importance
m = Wm
ratings in Table 1 and CorCR = [CorCRmn ]M×M be the correlation ma-
trix of the WHATs (CRs), then the new/adjusted matrix of importance
ratings can be calculated by
CRI
= CRI ⊗ CorCR (24)
where ⊗ is used to represent the linguistic aggregation operation.
k1 , W
k2 , . . . , W
kM ),
Since the order-based semantics of linguistic label is used in this
paper, the linguistically ordinal aggregation (Yager, 2007) may be a
good solution to Eq. (24).
Similarly, let R = [Rmn ]M×N be the matrix of the relationships
between WHATs (CRs) and HOWs (DRs) in Table 1 and CorDR =
[CorDRmn ]N×N be the correlation matrix of the HOWs (DRs), then the
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
[m5G;October 10, 2014;12:11]
new/adjusted matrix of the relationships between WHATs (CRs) and
HOWs (DRs) can be obtained by
R
(25)
where ⊗ is used to represent the linguistic aggregation operation,
which can be the linguistically ordinal aggregation (Yager, 2007).
With the new/adjusted matrices (CRI
text can take the correlations among WHATs and among HOWs into
account.
6.2. Planning matrix
The elements in the planning matrix consist of competitors,
customer competitive assessments, and goals for customer needs
(CRs/WHATs). Competitors who produce similar products may be
identified by the firm in a QFD context. Knowing the firm’s strengths
and constraints in all aspects of a product and in comparison with
its main competitors is essential for a firm if it wishes to improve its
competitiveness in the relevant markets. This kind of information can
be obtained by asking the customers to rate the relative performance
of the firm and its competitors on each WHAT (CR). Useful ways of
conducting this kind of comparison analysis are also via mailed sur-
veys and individual interviews. Denote the firm in question by F1 .
Suppose L − 1 competitors are identified, denoted as F2 , . . . , FL . Then
the K customers are requested to provide their subjective perceptions
on the relative performance of these L firms’ products performance
regarding the M WHATs (CRs) with a linguistic variable. Formally, the
judgment matrix of the relative performance of these L firms’ prod-
ucts regarding
k
by Pk = Pml , where k = 1, . . . , K.
M×L
Based on the order-based semantics of linguistic variable, we can
build M matrices of the L firms regarding
customer, denoted by PVmk = PVmk ln , where k = 1, . . . , K; m =
L×L
1, . . . , M. Then similar with Chan and Wu (2005) procedure, the per-
formance scores of firm 1 under customer k regarding customer need
m can be obtained by
1 L
xmk = PVmk
1l , (26)
L l=1
where m = 1, . . . , M; k = 1, . . . , K. Using the entropy method intro-
duced in Chan and Wu (2005), the firm 1 (F1 )’s priorities on the M
WHATs under customer k (Ck ) can be obtained and expressed by
Ek = (e1k , e2k , . . . , eMk ), where k = 1, . . . , K.
Performance goals on the WHATs (CRs) can be set by the firm
competitively and realistically, which is a highly strategical activity
involving many considerations from relevant management. Similar to
Chan and Wu (2005), the goals are assumed to numerical values and
normalized to [0, 1], denoted by GC = (gc1 , gc2 , . . . , gcM ). With the
performance scores of these L firms’ products performance regarding
the M WHATs and performance goals on the WHATs (CRs) of the firm
1 obtained, the new importance weight of customer need m with
respect to customer k can be calculated as follows
gcm
W
k k
× × emk (27)
xmk
k is the initial impor-
where k = 1, . . . , K; m = 1, . . . , M. The value Wm
tance weight of customer need m with respect to customer k de-
rived in Section 3, the value gcm /xmk indicates the firm’s improvement
ratio for customer need m with respect to customer Ck . With the new
weighting vector under customer k, denoted by (W
we are able to incorporate the planning matrix into our QFD context.
6.3. Technical matrix
This step is usually conducted to perform technical competitive
analysis through marketing. Although some technical parameters and
JID: EOR
ARTICLE IN PRESS
H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
know-hows of the competitors’ products cannot be easily obtained
and some may even be kept confidential, the producing firm should
make every effort to acquire this information and failing to do so may
result in an unfavorable position of the firm in the market place. In
case of extreme difficulty in obtaining the technical parameters of the
competitors’ products on some HOWs, careful technical assessments
should be made to give reliable scores with a linguistic variable repre-
senting the technical performance of the competitors’ products on the
said HOWs. Then the QFD team is requested to provide its subjective
perceptions on the relative performance of these L firms’ products
regarding the N HOWs (DRs) with a linguistic variable. Formally, the
linguistic judgment matrix of the technical parameter or performance
score of these L firms’ products regarding the N HOWs is denoted by
TP = [TPln ]L×N . Based on the order-based semantics of linguistic vari-
able, we can build N matrices of the L firms regarding the N HOWs,
[m5G;October 10, 2014;12:11]
13
Typically, the proposed QFD approach focuses on a simply but
widely used framework. Although our proposed approach is appeal-
ing and intuitively natural, some directions remain in which we must
extend our proposed approach before hopefully bringing it into prac-
tical application. Firstly, as summarized by Chan and Wu (2005), a
full HOQ contains as many as 18 elements. Our research framework
considers only six elements. We have provided some considerations
about other elements and suggest some ways to handle them in a po-
tentially more comprehensive HOQ model in Section 6. Therefore, one
possible extension is to explore the potentially more comprehensive
HOQ model based on our proposed approach in greater detail. Sec-
ondly, we only consider the first stage of the QFD. We hope our pro-
posed approach can be extended successfully to the entire four stages
of the QFD. Thirdly, as a similar methodology with QFD, Kansei Engi-
neering has been developed as a new product development method-

expressed as TPVn = TPVnml L×L where n = 1, . . . , N. Similar with the
planning matrix, we can also derive a technical performance score yn
of a HOW DRn as follows
1 L
yn = TPVn1l .
L l=1
Using the entropy method introduced in Chan and Wu (2005), the firm
F1 ’s priorities on the N HOWs can also be obtained and expressed by
Z = (z1 , z2 , . . . , zN ).
Finally, the new/adjusted priority weights for the N DRs can be
obtained by means of the following function
max{yn , gdn }
Vn = Vn × × zn , n = 1, . . . , N,
min{yn , gdn }
where Vn is the initial priority value derived by our approach, yn is
the technical performance score of firm F1 on HOW DRn , gdn is the
firm’s technical goal on HOW DRn , and zn is the priority value of firm
F1 on HOW DRn . The value max{yn , gdn }/min{yn , gdn } indicates the
technical improvement ratio (Chan & Wu, 2005). By this way, we are
able to include the technical matrix into our QFD context.
7. Concluding remarks
The use of QFD for new product design has proved to be a typical
group decision-making problem and requires the active participation
of customers and QFD design team. In order to achieve maximum
customer satisfaction with limited organizational resources such as
budget, it is essential to prioritize DRs (HOWs) reasonably so that lim-
ited resources can be optimally allocated among them. However, it is
not so straightforward for the QFD design team to prioritize the DRs
due to the following two types of uncertainties: one caused by human
perception and another due to customer heterogeneity in product
planning. Such two types of uncertainties may lead to the inaccurate
and wrong ranking of the DRs. In order to simultaneously address the
two types of uncertainties underlying QFD, a novel group decision-
making method is proposed in this paper. Firstly, the order-based
semantics of linguistic information is used to derive the individual
matrix of fuzzy preference relations of different DRs with respect to
each customer. Secondly, the concept of fuzzy majority underlying
group decision-making is used to perform the group aggregation of
individual fuzzy preference relations by means of an exponential RIM
quantifier. Thirdly, the fuzzy majority is also used to derive a consen-
sus degree for the uncertain QFD problem with the exponential RIM
quantifier. Two examples are borrowed from the literature to illus-
trate the proposed approach. It is shown that the proposed approach
outperforms the existing approaches (Chen et al., 2006; Kwong et al.,
2011; Wang, 2012) in terms of ease of use (the burden of quantify-
ing qualitative concepts can be eliminated), customer heterogeneity,
and the design team’s preference information α . Due to its easiness
in practice, our approach can reduce the cognitive burden of QFD
planning team and give a practical convenience in the process of QFD
planning.
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
ology to “translate the technology of a consumer’s feeling and image
for a product into the design elements of the product” (Nagamachi,
2002). There is a trend to incorporate the QFD into Kansei Engineering
(Hartono, Tan, Peacock & Ishihara, 2012). Consequently, we hope our
(28) proposed approach can be extended to Kansei Engineering.
Perhaps the biggest drawback to the proposed approach is the
requirement of an additional effort in deriving the linguistic quan-
tifier from decision-maker’s attitudinal preference, as shown in
Appendix A. However, the proposed bisection algorithm for deriving a
linguistic quantifier is computerized by Java language. Since develop-
ment of new information technologies (IT) appears to be revolution-
izing commerce generally and product development to a consider-
(29)
able degree (Krishnan & Ulrich, 2001), we believe that the proposed
approach with the help of IT will definitely reduce the complexity
and difficulty in practice. However, the benefit of new tools to man-
age product knowledge and support development decision making
within the extended enterprise needs to be explored in greater detail.
Acknowledgments
We would like to appreciate constructive comments and valu-
able suggestions from the three anonymous referees, which have
helped us efficiently improve the quality of this paper. This study
was partly supported by the National Natural Sciences Foundation of
China (NSFC) under grant nos. 71101050, 71125002, and 71471063;
sponsored by the Shanghai Pujiang Program and the Innovation Pro-
gram of Shanghai Municipal Education Commission under grant no.
14ZS060; and supported by the Fundamental Research Funds for the
Central Universities in China under grant no. WN1424004.
Appendix A. A RIM quantifier with exponential generating
function and the bisection search algorithm
Any relative quantifier can be expressed as a fuzzy subset Q of the
unit interval [0, 1]. In this representation for any proportion x ∈ [0, 1],
Q (x) indicates the degree to which x satisfies the concept conveyed
by the term Q. Yager (1996) further defined a Regular Increasing
Monotone (RIM) quantifier to represent the linguistic quantifier. A
fuzzy subset Q of the universe domain [0, 1] is called a RIM quantifier
if
Q (0) = 0, Q (1) = 1, and Q (x) ≥ Q (y) for x ≥ y.
Given a RIM quantifier Q, we can associate with this quantifier a
degree of orness (Yager, 1988), defined as
 1
orness(Q ) = Q (x)dx
0
= α (A.1)
which is equal to the area under a RIM quantifier. For f (x) in [0, 1] and
a RIM quantifier Q (x), f (x) is called the generating function of Q (x), if
JID: EOR
ARTICLE IN PRESS [m5G;October 10, 2014;12:11]

14 H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15

it satisfies to derive the membership function of the exponential RIM quanti-

 x fier with a given α . Toward this end, we have developed a bisection
Q (x) = f (t)dt, (A.2) algorithm to get the membership function of the exponential RIM
0
quantifier Q with a given α , as shown in Algorithm 1. It should be
1
where f (x) ≥ 0 and 0 f (x)dx = 1. It is obvious that, for any differen- noted that λ can theoretically be any value in (−∞, +∞). However, by
tiable RIM quantifier Q (x), its generating function exists, as f (x) can implementing the bisection algorithm with Java language, we found
be set as its first order differential function Q
(x) directly. With the that λ ∈ [−200, 200] is efficient enough to derive the membership
generation function f (x), the orness of Q (x) can be represented as function of the exponential RIM quantifier Q with any α ∈ [0, 1].
 1  x
α= f (t)dtdx Algorithm 1 A bisection search method to find Q (x) with an α value:
0

0
1
λ ∈ [−200, 200].
= 1− tf (t)dt (A.3) 1: if α = 0 then

0 0, if x = 1
2: Q (x) =
For two RIM quantifiers Q (x) and G(x) with their respective dif- 1, if x = 1
ferentiable and nonzero generating functions f (x) and g(x), if for all 3: else if α =1 then


0, if x = 0
x ∈ [0, 1], f (x), g(x) = 0 and ≤ f (x) g (x)
then orness(Q ) ≥ orness(G), 4: Q (x) =
f (x) g(x) , 1, if x = 0
refer to Liu (2006) for the proof. For a RIM quantifier with generating 5: else if α = 0.5 then
function f (x), if we want to make it more orlike (optimistic) or and- 6: Q (x) = x
like (pessimistic) we only need to increase or to decrease the value 7: else

of expression ff ((xx)) , respectively. The simplest case is to set ff ((xx)) = λ 8: if 0.5 < α < 1 then
as the control parameter of the orness level α for the RIM quantifier. 9: Initialize lower = −200, upper = 0, λ = −1
Therefore, it is easy to obtain the following function 10: else
11: Initialize lower = 0, upper = 200, λ = 1
f (x) = ceλx , 12: end if
1 while lower ≤ λ ≤ upper do
where c is a constant. Considering 0 f (x)dx = 1, we know c = λλ − 1. 13:
λ
e
Consequently, we will have the following expressions in terms of the 14: α
= eλ(e−λλ−1
−1
)
values α and λ (Liu, 2006): 15: if α
= α then
λx
⎧ 16: Q (x) = e λ −1
⎪ λeλx
⎨ f (x) = eλλx−1
e −1
⎪ 17: break
Q (x) = eeλ −1−1
(A.4) 18: else if α
> α then
⎪
⎪
⎩α= eλ −λ−1
19: lower ← λ
λ(eλ −1)
20: λ ← (lower + upper)/2
where λ ∈ (−∞, +∞) and 21: else
22: upper ← λ
• limλ→0 f (x) = 1, limλ→0 Q (x) = x, and limλ→0 α = 0.5, which indi- 23: λ ← (lower + upper)/2
cates that f (x), Q (x), α can be seen as continuous at λ = 0. The case 24: end if
of λ = 0 indicates that the decision-maker has a neutral attitude, 25: end while
the membership function of Q is Q (x) = x. 26: end if
• If α < 0.5, then λ > 0, f (x) is increasing, the decision-maker can 27: return Q (x)
be seen as pessimistic. The more closer α to 0, the bigger λ is, the
more increasing extent of f (x) will be, and α will become smaller.
When α → 0, then λ → +∞ and

0, if x = 1 References
Q (x) →
1, if x = 1
Chan, L.-K., Wu, M.-L. (2002a). Quality function deployment: A comprehensive review
• If α > 0.5, then λ < 0, f (x) is decreasing, the decision-maker can of its concepts and methods. Quality Engineering, 15, 23–35.
Chan, L.-K., Wu, M.-L. (2002b). Quality function deployment: A literature review. Euro-
be seen as optimistic. The more closer α to 1, the smaller λ is, the pean Journal of Operational Research, 143, 463–497.
more decreasing extent of f (x) will be, and α will become smaller. Chan, L.-K., Wu, M.-L. (2005). A systematic approach to quality function deployment
When α → 1, then λ → −∞ and with a full illustrative example. Omega, 33, 119–139.
 Chen, L.-H., Ko, W.-C. (2010). Fuzzy linear programming models for NPD using a four-
0, if x = 0 phase QFD activity process based on the means-end chain concept. European Journal
Q (x) = of Operational Research, 201, 619–632.
1, if x = 0 Chen, L.-H., Ko, W.-C. (2011). Fuzzy nonlinear models for new product development us-
• The orness value (α ) of a RIM quantifier Q with exponential gen- ing four-phase quality function deployment processes. IEEE Transactions on Systems,
Man, and Cybernetics—Part A: Systems and Humans, 41, 927–945.
erating function, is strict monotonic decrease with parameter λ Chen, L.-H., Ko, W.-C., Tseng, C.-Y. (2013). Fuzzy approaches for constructing house of
such that ∂α
∂λ < 0.
quality in QFD and its applications: A group decision-making method. IEEE Trans-
actions on Engineering Management, 60, 77–87.
Chen, L.-H., Weng, M.-C. (2006). An evaluation approach to engineering design in qfd
From the exponential generating function and exponential RIM
processes using fuzzy goal programming models. European Journal of Operational
quantifier introduced in Eq. (A.4), it is found that different mem- Research, 172, 230–248.
bership functions of an exponential RIM quantifier will be obtained Chen, Y., Fung, R. Y., Tang, J. (2006). Rating technical attributes in fuzzy QFD by integrat-
according to the value α . Also, in order to get the membership func- ing fuzzy weighted average method and fuzzy expected value operator. European
Journal of Operational Research, 174, 1553–1566.
tion of a RIM quantifier with a given α , we need to calculate the value Cristiano, J. J., Liker, J. K., WhiteIII, C. C. (2001). Key factors in the successful application
λ associated with the value α . The membership functions of the ex- of quality function deployment (QFD). IEEE Transactions on Engineering Management,
ponential RIM quantifier with respect to α = 0, 0.5, 1 are consistent 48, 81–94.
Ertay, T., Kahraman, G. C., Ruan, D. (2005). Quality function deployment implemen-
with the ones derived by traditional approaches (Yager, 1988). How- tation based on analytic network process with linguistic data: An application in
ever, when α ∈ (0, 0.5) and α ∈ (0.5, 1), it is not so straightforward automotive industry. Journal of Intelligent & Fuzzy Systems, 16, 221–232.

Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
JID: EOR
ARTICLE IN PRESS
H.-B. Yan, T. Ma / European Journal of Operational Research 000 (2014) 1–15
Han, C.-H., Kim, J.-K., Choi, S. H. (2004). Prioritizing engineering characteristics in qual-
ity function deployment with incomplete information: A linear partial ordering
approach. International Journal of Production Economics, 91, 235–249.
Hartono, M., Tan, K. C., Peacock, J. B., Ishihara, S. (2012). Incorporating Markov chain
modelling and QFD into Kansei engineering applied to services. International Journal
of Human Factors and Ergonomic, 1, 74–97.
Hauser, J. R., Clausing, D. (1988). The house of quality. Harvard Business Review, 66,
63–73.
Herrera, F., Martínez, L. (2000). A 2-tuple fuzzy linguistic representation model for
computing with words. IEEE Transactions on Fuzzy Systems, 8, 746–752.
Ho, E. S. S. A., Lai, Y.-J., Chang, S. I. (1999). An integrated group decision-making approach
to quality function deployment. IIE Transactions, 31, 553–567.
Kacprzyk, J., Zadrozny, S., Fedrizzi, M., Nurmi, H. (2008). On group decision making,
consensus reaching, voting and voting paradoxes under fuzzy preferences and a
fuzzy majority: A survey and some perspectives, In: H., Bustince F., Herrera J.,
Montero (Eds.), Fuzzy sets and their extensions: Representation, aggregation and mod-
els (pp. 263–295). Heidelberg: Physica-Verlag.
Kao, C., Liu, S. T. (2001). Fractional programming approach to fuzzy weighted average.
Fuzzy Sets and Systems, 120, 435–444.
Karsak, E. E. (2004). Fuzzy multiple objective decision making approach to prioritize
design requirements in quality function deployment. International Journal of Pro-
duction Research, 42, 3957–3974.
Khoo, L. P., Ho, N. C. (1996). Framework of a fuzzy quality function deployment system.
International Journal of Production Research, 34, 299–311.
Kim, K.-J., Moskowitz, H., Dhingra, A., Evans, G. (2000). Fuzzy multicriteria models
for quality function deployment. European Journal of Operational Research, 121,
504–518.
Krishnan, V., Ulrich, K. T. (2001). Product development decisions: A review of the
literature. Management Science, 47, 1–21.
Kwong, C., Chen, Y., Bai, H., Chan, D. (2007). A methodology of determining aggregated
importance of engineering characteristics in QFD. Computers & Industrial Engineer-
ing, 53, 667–679.
Kwong, C. K., Ye, Y., Chen, Y., Choy, K. L. (2011). A novel fuzzy group decision making
approach to prioritising engineering characteristics in QFD under uncertainties.
International Journal of Production Research, 49, 5801–5820.
Liou, T. S., Wang, M. J. (1992). Fuzzy weighted average: An improved algorithm. Fuzzy
Sets and Systems, 49, 307–315.
Liu, B., Liu, K. L. (2002). Expected value of fuzzy variable and fuzzy expected value
models. IEEE Transactions on Fuzzy Systems, 10, 445–450.
Liu, C.-H., Wu, H.-H. (2008). A fuzzy group decision-making approach in quality function
deployment. Quality & Quantity, 42, 527–540.
Liu, S.-T. (2005). Rating design requirements in fuzzy quality function deployment via a
mathematical programming approach. International Journal of Production Research,
43, 497–513.
Please cite this article as: H.-B. Yan, T. Ma, A group decision-making approach to uncertain quality function deployment based on fuzzy
preference relation and fuzzy majority, European Journal of Operational Research (2014), http://dx.doi.org/10.1016/j.ejor.2014.09.017
[m5G;October 10, 2014;12:11]
15
Liu, X. (2006). Some properties of the weighted OWA operator. IEEE Transactions on
Systems, Man, and Cybernetics–Part B: Cybernetics, 36, 118–127.
Nagamachi, M. (2002). Kansei Engineering as a powerful consumer-oriented technol-
ogy for product development. Applied Ergonomics, 33, 289–294.
Saaty, T. L. (1986). Axiomatic foundation of the analytic hierarchy process. Management
Science, 32, 841–855.
Shen, X. X., Tan, K. C., Xie, M. (2001). The implementation of quality function deploy-
ment based on linguistic data. Journal of Intelligent Manufacturing, 12, 65–75.
Temponi, C., Yen, J., Tiao, W. A. (1999). House of quality: A fuzzy logic-
based requirements analysis. European Journal of Operational Research, 117,
340–354.
Tidd, J., Bodley, K. (2002). The influence of project novelty on the new product devel-
opment process. R&D Management, 32, 127–138.
Vanegas, L. V., Labib, A. W. (2001). A fuzzy quality function deployment (FQFD)
model for deriving optimum targets. International Journal of Production Research, 39,
99–120.
Wang, J. (1999). Fuzzy outranking approach to prioritize design requirements in
quality function deployment. International Journal of Production Research, 37,
899–916.
Wang, Y.-M. (2012). A fuzzy-normalisation-based group decision mak-
ing approach for prioritising engineering design requirements in QFD
under uncertainty. International Journal of Production Research, 50,
6963–6977.
Wasserman, G. S. (1993). On how to prioritize design requirements during the QFD
planning process. IIE Transactions, 25, 59–65.
Xu, Y., Da, Q., Liu, L. (2009). Normalizing rank aggregation method for priority of a
fuzzy preference relation and its effectiveness. International Journal of Approximate
Reasoning, 50, 1287–1297.
Yager, R. R. (1988). On ordered weighted averaging operators in multi-criteria decision
making. IEEE Transactions on Systems, Man and Cybernetics, 18, 183–190.
Yager, R. R. (1996). Quantifier guided aggregation using OWA operators. International
Journal of Intelligent Systems, 11, 49–73.
Yager, R. R. (2007). Aggregation of ordinal information. Fuzzy Optimization Decision
Making, 6, 199–219.
Yan, H.-B., Ma, T., Li, Y. (2013). A novel fuzzy linguistic model for prioritising engi-
neering design requirements in quality function deployment under uncertainties.
International Journal of Production Research, 51, 6336–6355.
Zadeh, L. (1983). A computational approach to fuzzy quantifiers in natural languages.
Computers and Mathematics with Applications, 9, 149–184.
Zadeh, L. A. (1965). Fuzzy sets. Information and Control, 8, 338–353.
Zadeh, L. A. (1975). The concept of a linguistic variable and its application to approxi-
mate reasoning—Part I. Information Sciences, 8, 199–249.
Zhou, M. (1998). Fuzzy logic and optimization models for implementing QFD. Computers
and Industrial Engineering, 35, 237–240.