Professional Documents
Culture Documents
2020 - Adaptive Consensus Reaching Process With Hybrid Strategies For Large-Scale Group Decision Making
2020 - Adaptive Consensus Reaching Process With Hybrid Strategies For Large-Scale Group Decision Making
Decision Support
a r t i c l e i n f o a b s t r a c t
Article history: Large-scale group decision making, which involves dozens to hundreds of experts, is attracting increasing
Received 4 December 2018 attention and has become an important topic in the field of decision making. Because of the clustering
Accepted 2 October 2019
process, a large-scale group decision making problem can be divided into two levels: inter sub-group and
Available online 10 October 2019
intra sub-group. In existing consensus models under the large-scale group decision making environment,
Keywords: the degree of consensus within the intra sub-group is not truly taken into account. To deal with this
Decision support systems issue, this work develops an adaptive consensus model for the sub-groups composed of hybrid strategies,
Large-scale group decision making with or without a feedback mechanism, according to the different levels of inter and intra degrees of
Adaptive consensus consensus. These different levels of consensus are divided into four scenarios (high–high, high–low, low–
Hybrid strategies high, low–low), and different feedback suggestions are generated corresponding to different cases. This
Reciprocal comparison matrices hybrid mechanism can reduce the cost of supervision for the moderator. The fuzzy c-means clustering
algorithm is used to classify experts. A weight-determining method combining the degree of cohesion and
the size of a sub-group is introduced. Finally, an illustrative example is offered to verify the practicability
of the proposed model. Some discussions and comparisons are provided to reveal the advantages and
features of the proposed model.
© 2019 Elsevier B.V. All rights reserved.
https://doi.org/10.1016/j.ejor.2019.10.006
0377-2217/© 2019 Elsevier B.V. All rights reserved.
958 M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971
Gupta, 2018; Mata, Martínez & Herrera-Viedma, 2009; Pang, Liang Table 1
List of notations.
& Song, 2017; Pérez, Cabrerizo, Alonso & Herrera-Viedma, 2014;
Rodríguez, Labella, Tré & Martínez, 2018; Zhang, Zhu, Liu, Chen & Notations Meanings
Ma, 2017). Among these models, the adaptive CRP can generate X = {x1 , x2 , . . . , xn } A set of n alternatives
recommendations adaptively by considering dynamic parameters E = {e1 , e2 , . . . , em } A set of m experts
such as levels of group consensus (Mata et al., 2009) and the P k = ( pki j )n×n The reciprocal comparison matrix (RCM) of the kth expert
weights of experts (Pérez et al., 2014). For instance, Mata et al. {C 1 , C 2 , . . . , C N } N clusters
ch Cluster center
(2009) proposed a consensus model in which the amount of
μch,t (Pk ) Membership degree of P k to the selected centroid ch,t
recommendations and the number of experts involved in each d (P k , ch,t ) Distance between an RCM P k and the cluster center ch,t
iteration were adapted to different levels of group consensus. At #C h,t Number of experts in cluster C h,t
the initial iteration stage, given that the level of consensus is s(C h ) Cohesion degree of cluster C h
very low, all the experts are advised to modify their evaluation ω (C h ) Normalized priority vector of cluster C h
CL(P k ) Consensus level of expert ek
information. With the level of consensus increases, both the num- SCL Consensus level of a sub-group
ber of changes and the number of experts who need to modify SCL Threshold of SCL
their evaluations decrease. This adaptive CRP was based on four ICL Inter consensus level of a sub-group to the global group
categories of consensus levels: very low, low, medium, and high. GCL Global consensus level
GCL Threshold of GCL
Pérez et al. (2014) introduced a CPR that adapted to the weights
GSC Global/sub-group consensus
of experts. The logic of this model is that the experts with higher c pi j Degree of consensus on a position
weights have greater expertise, and thus, they can be asked to c ai Degree of consensus on an alternative
perform a smaller amount of modifications in the CRP.
The existing methods can be used to handle GDM problems
under different circumstances. However, there are still some
limitations: strategies with or without a feedback mechanism according to the
different degrees of inter and intra agreements, is presented in
(1) The clustering process plays a vital role in solving LSGDM detail. Finally, a case study is provided to better understand the
problems, and this is the main difference between general theory of this study.
GDM and LSGDM. Determining a cluster’s weight is critical The primary contributions of this work are summarized as
in LSGDM. Most of the existing studies (Wu & Xu, 2018; Xu, follows:
Du et al., 2015; Zhang et al., 2018) simply used the num-
ber of experts in a cluster as the standard to determine the
(1) A weight-determining method for clusters is developed. This
weights of clusters. In this way, the clusters with the same
method considers both the size and the cohesion degree of
number of experts but possessing different inner character-
a cluster, and thus improves the weight-determining method
istics in terms of cohesiveness or diversity would still have
that only considers the size of a cluster. Furthermore, a co-
the same weights. Thus, we need to develop a method to de-
hesive cluster will exert much influence in the global group,
termine the weights of different clusters that considers both
and thus this method can accelerate the convergence speed
the degree of cohesion and the size of a cluster.
of the CRP objectively since the aim of the CRP is to make
(2) The use of the adaptive consensus model in an LSGDM con-
the group information central. An aggregation formula is
text is still at an initial exploratory stage, and thus, it faces
proposed to integrate these two parameters. The degree of
big challenges. In LSGDM, owing to the various backgrounds
cohesion of a cluster includes the silhouette coefficient.
and expertise of experts, it is more difficult to reach a con-
(2) According to the different levels of inter and intra consensus
sensus result than in conventional GDM. Thus, it is necessary
of a sub-group, this study develops an adaptive consensus
to research the adaptive consensus model within a large-
model with hybrid strategies for multiple groups. This model
scale context. Although Rodríguez et al. (2018) introduced an
can generate different recommendations according to the
adaptive consensus model under the LSGDM context similar
degrees of consensus of two levels. These different strategies
to the method of Mata et al. (2009), their model classified
can reduce the numbers of involved experts and the recom-
the global consensus into three levels: low, high, and high
mendations for modifications in the CRP within the LSDGM
enough. Nevertheless, this model did not take into account
context. Thus, the proposed method can reduce the supervi-
the intra consensus of a sub-group. In LSGDM, the cluster-
sion and adjustment costs of the CRP.
ing process should be used to cluster the experts into sev-
(3) An illustrative example about the legislative amendment of
eral groups based on their preference information to sim-
the International Trade Law within the United Nations sys-
plify the decision-making process. As a result, two levels of
tem is provided to explore and verify the applicability and
consensus are generated, i.e., the level of consensus within
feasibility of the proposed model. Then, comparisons regard-
a sub-group and the level of consensus of a sub-group to
ing different parameters including the number of clusters
the global group. These two kinds of consensus are respec-
and the consensus threshold are provided. We also give a
tively named as the intra consensus and inter consensus in
comparison with Rodríguez et al. (2018)’s method to validate
this study. The existing studies about LSGDM did not gen-
the advantage of the adaptive mechanism.
erate feedback suggestions considering these two kinds of
consensus at the same time.
To achieve the above goals, this paper is organized as follows:
To address these shortcomings, in this study, we develop a Section 2 describes some concepts used in this study, including
novel adaptive consensus method under the LSGDM context. the LSGDM, CRP, and FCM. Section 3 presents a novel adaptive
This method first uses the fuzzy c-means (FCM) clustering al- consensus model under the LSGDM context. In Section 4, an
gorithm to classify the experts into several clusters. Then, a illustrative example is included to verify the applicability and
weight-determining method for clusters is proposed. Afterwards, feasibility of this model. Some comparisons and discussions are
the consensus measures corresponding to the intra consensus and also provided. Concluding remarks are presented in Section 5.
inter consensus are defined. Based on these consensus measures, To facilitate the comprehension of the paper, Table 1 summa-
an adaptive feedback mechanism, composed of hybrid consensus rizes some of the used notations.
M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971 959
Moderator
Moderator
Problem
Problem Experts
Experts Feedback
Feedback recommendations
recomm
mendations
Low
ow consensus
consensus
level
level High
High consensus
consensus
Alternatives
Altern
r atives level
level
Consensus
Consensuus
control
contro
ol
Selection
Selection
Preference
Prefe
f rence Calculating
alculating process
process
relations
relations consensus
us degree
consensu degreee
1) Collecting preferences. The preferences evaluated by experts are 1) The fuzziness parameter b is the weighting exponent in
collected. FCM to control the extent of sharing among clusters
960 M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971
6) The FCM algorithm executes the termination condition when all 3.3. Computing the degrees of consensus
clusters are stabilized. This occurs when the difference between
membership degrees in two successive iterations approaches to Expert ek provides preferences on a set of alternatives by an
0, i.e., RCM P k = ( pki j )n×n with n(n − 1 )/2 pairwise comparisons. Then, a
m N moderator computes the degrees of consensus of the expert group
k=1 h=1 |μch,t+1 (Pk ) − μch,t (Pk )|
≤ε by distance measures.
m·N Let P k = ( pki j )n×n be an RCM provided by expert ek , and
where ε is a parameter whose value approaches 0. In this work, P c = ( pci j )n×n be the group preference relation, which can be
we set ε as 0.001. represented by its centroid ci j . Then, the consensus level (CL) of
expert ek is defined as:
3.2. Determining the weights of clusters
CL(P k ) = 1 − d (P k , P c ) (3)
After classifying the experts into multiple sub-groups, the next where d (P k , P c ) is a distance measured between P k and P c .
task is to assign weights to each cluster. One principle is that The consensus level of a sub-group C h (SC L(C h )) is computed
the larger cluster with more experts should be assigned a larger by
weight. Another indicator that should be considered is the degree
1
of cohesion of a cluster. The higher the degree of cohesion of SC L(C h ) = C L (P k )
a cluster is, the larger the weight that should be given to this #C h k h
P ∈C
cluster. To measure the degree of cohesion of a cluster, we use the
silhouette coefficient introduced by Rousseeuw (1987) and adapt where #C h
is the number of experts in the sub-group C h .
it to the FCM algorithm. If SC L(C h ) = 1, all experts in the sub-group C h have a unified
Let A(P k ) be the average dissimilarity of an RCM P k to all other opinion. A larger value of SC L(C h ) represents a higher consensus
data objects in cluster C a . Dissimilarity is a type of proximity level among the experts in the cluster. SCL is a predefined thresh-
that measures how far away two objects are from each other old for SC L(C h ). If SC L(C h ) ≥ SCL, then the sub-group has an ac-
(Rousseeuw, 1987). It is contrary to another widely used type of ceptable consensus; otherwise, the CRP should be applied to reach
proximity: similarity. Here, we use the Euclidean distance to define an acceptable degree of consensus. Note that SCL should be set
the dissimilarity degree. Let d (P k , C h ) be the average dissimilarity in advance, which represents the expected standard that a group
degree of P k to all data objects in cluster C h , h = a. Then, the needs to achieve in LSGDM. In a real-world decision-making prob-
smallest value of d (P k , C h ) is selected as B(P k ), where lem, a high value of SCL may be difficult to achieve or needs many
iterations. Thus, we use a parameter (max _round) to set the max-
B(P k ) = min d (P k , C h ), P k ∈ C a imum number of iterations. The value of SCL can be determined
h=a
according to specific decision-making problems. In classical GDM
The silhouette coefficient of an RCM P k is obtained by: problems, the value of the predefined threshold is usually set as
0.9 (Tian, Zhang & Ha, 2018), 0.85 (Rodríguez et al., 2018), 0.8 (Wu
1 − A(P k )/B(P k ), if A(P k ) < B(P k ) & Xu, 2016), or 0.75 (Herrera-Viedma, Martínez & Mata, 2005). For
s (P ) =
k
0, if A(P k ) = B(P k ) some vital and major issues, the minimum value of consensus level
B(P k )/A(P k ) − 1, if A(P k ) > B(P k ) should be high (Herrera-Viedma et al., 2005). However, in other
or extreme cases, such as emergency problems, the consensus thresh-
old should be low (Xu, Du et al., 2015) because of the time limit.
B (P k ) − A (P k )
s (P k ) = In LSGDM problems, a low threshold is also needed because there
max{A(P k ), B(P k )} are great differences in experts’ background, knowledge, and ex-
It is easy to find that −1 ≤ s(P k ) ≤ 1. The closer the value of pertise. For the reason that SCL is the consensus threshold within a
s (P k )
is to 1, the better P k attached to its cluster should be. sub-group, in this study, we set SCL as 0.9, a relatively high value.
To measure the degree of cohesion of a sub-group, we define The inter consensus level (IC L(C h )) between a cluster C h and
the silhouette coefficient of a cluster in the FCM as: the global preference relation P G = ( pGi j )n×n is defined as:
1 IC L(C h ) = 1 − d (P G , C h ) (4)
s(C h ) = s (P k )
#C h k h
P ∈C where PG
is obtained by aggregating all sub-groups’ preference
relations, such that
Given that s(C h )
can be used to measure the degree of cohesion
of a cluster C h and the number of experts in the cluster C h , #C h ,
N
can also be used to reflect the importance of the cluster, we pGi j = ω (C u ) · pui j
aggregate these two indicators as: u=1
consensus at the alternative level as γ S−HL , such that ii) The positions P OSS−LH to be modified are identified as
n
P OSS−LH = {(i, j )|c pi j < γ S−LH ∧ xi ∈ ALT S−LH } (11)
γ S−HL = cai /n (8)
i=1 Generation of advice. Once the positions that need to be mod-
ified are identified, the direction rules that are used to give
iii) For any alternative xi ∈ ALT S−HL , we then identify the positions adjustment suggestions should be provided by the moderator. The
to be modified: directions rules are recommended as follows:
P OSS−HL = {(i, j )|c pi j < γ S−HL ∧ xi ∈ ALT S−HL } (9)
a) DR. 21: If cihj < pGi j , then all experts in the sub-group Ch
Here, γ S−HL denotes the average degree of consensus should increase the assessments associated with pairs of al-
at the position level. Because the degree of consensus at ternatives (xi , x j ).
an alternative level cai is calculated based on all positions b) DR. 22: If cihj > pGi j , then all experts in the sub-group Ch
associated with the alternative xi (see Eq. (5)), we have
n n n should decrease the assessments associated with pairs of al-
i=1 ( j=1,i= j c pi j )/ (n − n ) =
2
i=1 cai /n in the same prefer- ternatives (xi , x j ).
ence relation. We use the same symbol to represent the average
degree of consensus at the position level as that in Eq. (8). 4) Strategy for low–low consensus (S-LL). If the SC L(C h ) of a sub-
Generation of advice. This strategy provides suggestions for each ex- group C h does not reach the threshold SCL, nor does its level of
pert in a sub-group. The direction rules are described as follows: consensus to the whole group IC L(C h ), then this sub-group is
diverse and its centroid is far from the global collective point.
a) DR. 31: If pki j < cihj , then expert ek should increase the assess-
In this scenario, the moderator needs to provide suggestions to
ments associated with pairs of alternatives (xi , x j ); make the sub-group’s centroid converge to the global collective
b) DR. 32: If pki j > cihj , then expert ek should decrease the assess- point and also make the sub-group coherent. It seems reason-
ments associated with pairs of alternatives (xi , x j ). able to use both the global collective point and the cluster’s
center point to control the adjustment directions. The graphic
3) Strategy for low–high consensus (S-LH). If the SC L(C h ) of a sub- display for S-LL is included in Fig. 5. To do this, we define a
group C h reaches the threshold SCL while its inter consensus new measure, called the global/sub-group consensus (GSC),
level IC L(C h ) does not, then this sub-group is cohesive but its which is achieved by a control parameter δ :
centroid is far from the global collective point. In this situation,
there is no need to provide personalized recommendations for GSC = (1 − δ )CL(P k ) + δ IC L(C h ) (12)
all of the experts in the sub-group. To reduce supervision and where CL(P k ) is the degree of consensus of expert to ek
adjustment costs, it is appropriate to provide a unified sugges- the sub-group’s centroid, which can be obtained by Eq. (3);
tion for the sub-group. Then, this sub-group with its centroid IC LG (C h ) is the degree of consensus of sub-group C h to the
can converge toward the global center point. The graphic dis- global collective point, which can be obtained by Eq. (4).
play for S-LH is included in Fig. 4. δ ∈ [0, 1] is a parameter to control the weight of two levels
of consensus. Usually δ > 0.5 can be used to assign more
To find out the set of preferences that should be changed by
importance to the global consensus level.
each sub-group, C h ∈ S−LH, this strategy is developed as follows:
To find out the set of preferences that should be changed
i) In this scenario, it is not necessary to identify the experts who by experts in a sub-group, C h ∈ S − LL, the following strategy
contribute less in reaching a high level of consensus. First, the should be implemented:
alternatives ALT S−LH that have a smaller degree of consensus
than the predefined threshold γ S−LH are identified as i) Initially, we identify the experts, EX P S−LL , who should make
modifications in a sub-group. It is computed by
ALT S−LH = {xi |cai < γ S−LH } (10)
EX P S−LL = {ek |(1 − δ )C LSub (P k ) + δ IC LG (C h ) < (1 − δ )SCL
It is noted that the symbol γ S−LH
has the same calculation as
that in Eq. (8). The difference is that γ S−LH here is on the level
+ δ GCL} (13)
of a sub-group to the global group, while the former is on the The consensus threshold here is dynamic. It takes into account
level of intra sub-group. two levels of consensus at the same time.
964 M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971
Algorithm 2 Table 2
Adaptive CRP with hybrid strategies for LSGDM with RCMs. Weights of clusters.
Input: The original individual RCMs P k = ( pki j )n×n , the given consensus Cluster Degree of cohesion Size Weight
thresholds, SCL and GCL, and the parameter δ . C1 0.6286 6 0.3422
Output: The ranking of alternatives. C2 0.4700 3 0.1460
Step 1: Classify experts into several clusters {C 1 , C 2 , . . . , C N } via the FCM C3 0.4560 6 0.2879
algorithm. C4 0.3867 5 0.2239
Step 2: Determine the weights of clusters by Eq. (2).
Step 3: Calculate the levels of consensus of the two levels, SCL and GCL.
Step 4: If GCL ≥ GCL, then go to the selection process; otherwise, identify
sub-groups’ categories and go to the next step.
Step 5: Generate hybrid feedback adjustment rules to help experts modify 4.1. Case description
their preferences.
Step 6: Derive priority vector from the global collective RCM and rank the The United Nations Commission on International Trade Law
alternatives (UNCITRAL) is the only core legal institute in the field of inter-
Step 7: End.
national trade law within the United Nations system. Since the
end of the 20th century, trade tensions between countries around
the world are increasing, and trade disputes have intensified.
ii) For any expert ek ∈ EX P S−LL , this step identifies the alternatives UNCITRAL has made a great contribution for the unification of leg-
that should be modified: islation on international trade law. However, with the rapid devel-
opment of economic globalization, a growing number of countries
ALT S−LL = {xi |(1 − δ )caSub
i + δ caGi < γ S−LL ∧ ek ∈ EX P S−LL } (14)
are involved. This makes the development of the international
where γ S−LL = ( (1 − δ ) ni=1 caSub
i
+ δ ni=1 caGi )/n. trade law more diversified and the conflicts between different
iii) Finally, for any alternative xi ∈ ALT S−LL , this step identifies the legal systems more acute. The task of unification of legislation on
positions to be modified: international trade is facing big challenges and many problems
have arisen, such as fuzzy description, delayed legislative tools,
P OSS−LL = {(i, j )|(1 − δ )cpSub
i j + δ cpi j < γ
G S−LL
∧ xi ∈ ALT S−LL } and vague terms. Therefore, UNCITRAL has taken some measures
(15) to solve these problems. One way is to speed up the process of
democratic legislation. UNCITRAL absorbs recommendations from
Generation of advice. In this step, the moderator provides per- countries all over the world in legislative planning, text drafting,
sonalized recommendations to help the experts to modify their and draft reviewing. The main forms of democratic legislation
preferences. The new preference value on a pair of alternatives are international academic conferences, specialist symposiums,
(xi , x j ) that should be modified by the expert is generated as feasibility study meetings, and non-government organizations.
Suppose that four problems existing in old laws need to be
r pi j = (1 − δ )ciSub
j + δ pi j
G
(16) amended:
Finally, the direction rules are implemented as follows:
x1 : Unified possibility of digital currency rules;
a) DR. 1: If pki j < r pki j , then expert ek should increase the assess- x2 : Unification of drug cross-border trading rules;
ments associated with pairs of alternatives (xi , x j ); x3 : Feasibility of unification of the International Trade Contract
b) DR. 2: If pki j > r pki j , then expert ek should decrease the assess- Law;
ments associated with pairs of alternatives (xi , x j ). x4 : Legal principle of the confidentiality characteristics of Inter-
national Commercial Arbitration.
Note. The value of δ can reflect different degrees of emphasis
concerning the consensus of the two levels. As mentioned be- To select the alternative that needs to be solved firstly, UNCI-
fore, when δ > 0.5, the consensus model leads experts toward the TRAL invites 20 experts E = {e1 , e2 , . . . , e20 } from 20 different
global collective point rather than toward the sub-group’s centroid. countries to evaluate these alternatives. The experts provide their
If δ = 0, then GSC = CL(P k ). In this case, the global collective point preferences by RCMs. To save space, we put these 20 RCMs in the
is not used to guide experts to modify their preferences. This case Appendix.
goes into the third strategy S-HL. If δ = 1, then GSC = IC L(C h ) and As mentioned before, for vital and major issues, the consensus
the cluster center point is not used to guide experts to modify level should be high. Therefore, in this case, the parameters are
their preferences. This case goes into the second strategy S-LH. established as GCI = 0.8, SCI = 0.9, and δ = 0.7.
Therefore, the second and third scenarios can be regarded as two
extreme cases of the fourth scenario. Furthermore, the parameter
δ can also work when generating advice in the fourth scenario
4.2. Solving the problem by the large-scale adaptive CRP with hybrid
(see Eq. (16)). Different values of δ can reflect different emphases.
strategies
3.5. An adaptive CRP with hybrid strategies for LSGDM with RCMs To obtain the problem’s final solution, the proposed large-scale
adaptive CRP with hybrid strategies is applied.
Based on the above analyses, we provide the following algo- First round.
rithm for LSGDM with RCMs. The flow chart of the algorithm is
illustrated intuitively in Fig. 6. Step 1 Using the FCM algorithm, the 20 experts are clas-
sified into 4 sub-groups: C 1 = {e1 , e2 , e3 , e4 , e5 , e6 },
4. Illustrative example C 2 = {e7 , e8 , e9 }, C 3 = {e10 , e11 , e12 , e13 , e14 , e15 } and
C = {e , e , e , e , e20 }.
4 16 18 17 19
This section presents an illustrative example to verify the Step 2 Eq. (2) is used to obtain the weight of each cluster. Table
applicability of the proposed adaptive CRP with hybrid strategies 2 shows the degree of cohesion, size, and weight of each
for LSGDM problems with RCMs. cluster.
M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971 965
Start
Cohesion Size
NO
Finish
⎧ ⎫
Step 2 The RCMs of the four subgroups and the global group are: ⎪
⎨0.50 0 0 0.4667 0.1667 0.2167 ⎪
⎬
0.5333 0.50 0 0 0.2667 0.2833
PC2 = ,
⎩ 0.8333
⎪ 0.7333 0.50 0 0 0.5333 ⎪
⎭
0.7833 0.7167 0.4667 0.50 0 0
⎧ ⎫ ⎧ ⎫
⎪
⎨0.50 0 0 0.6833 0.8083 0.7500⎪
⎬ ⎪
⎨0.50 0 0 0.2917 0.3417 0.5333 ⎪
⎬
0.3167 0.50 0 0 0.7167 0.6417 0.7083 0.50 0 0 0.5750 0.8083
PC1 = , PC3 = ,
⎩ 0.1917
⎪ 0.2833 0.50 0 0 0.4083⎪
⎭ ⎩ .6583
⎪ 0 0.4250 0.50 0 0 0.6500⎪
⎭
0.2500 0.3583 0.5917 0.50 0 0 0.4667 0.1917 0.3500 0.50 0 0
966 M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971
Table 3
Degrees of consensus of the two levels.
GCL: 0.7889
C1 C2 C3 C4
⎧ ⎫
⎪
⎨0.50 0 0 0.7900 0.7400 0.4200⎪
⎬
Table 4
Identification rules for sub-group C 4 .
0.2100 0.50 0 0 0.4400 0.1800
PC4 = ,
⎪ 0.2600 0.5600 0.50 0 0 0.2300⎪ γ S−LL ALT S−LL POSS−LL
⎩ ⎭
0.5800 0.8200 0.7700 0.50 0 0 e16
0.8388 x2 , x4 (2,1), (2,4), (4,2)
⎧ ⎫ e17 0.8408 x4 (4,2), (4,3)
⎪
⎨0.50 0 0 0.5600 0.5613 0.5334 ⎪
⎬ e18 0.8048 x1 , x2 , x3 , x4 (1,2), (1,3), (2,1), (2,4), (3,2), (3,4), (4,2), (4,3)
0.4400 0.50 0 0 0.5465 0.5343 e19 0.8358 x2 , x4 (2,1), (2,4), (4,2)
PG =
⎩0.4387
⎪ 0.4535 0.50 0 0 0.4577 ⎪
⎭ e20 0.8218 x2 , x4 (2,1), (2,4), (4,2), (4,3)
0.4666 0.4657 0.5423 0.50 0 0
Step 3 Compute the level of consensus of the two levels. Table 3
The experts in sub-group C 4 are required to modify their
presents the degrees of consensus in the first round.
preference assessments as follows:
Step 4 Given that GCL = 0.7888 < 0.8, the feedback mechanism
Expert e16 should increase his assessments for positions (2,1)
should be applied. These four sub-groups’ categories are
and (2,4), and decrease his assessment for position (4,2).
identified as:
Expert e17 should decrease his assessments for positions (4,2)
Step 4 IC L(C 1 ) = 0.8358 > 0.8, SC L(C 1 ) = 0.9370 > 0.9, then C 1 ∈
and (4,3).
S − HH;
Expert e18 should increase his assessment for position (1, 2 ),
Step 4 IC L(C 2 ) = 0.7361 < 0.8, SC L(C 2 ) = 0.9044 > 0.9, then C 2 ∈
(2,4), (3,2), (3,4), and (4,2), and decrease his assessments for
S − LH;
positions (1,3), (2,1), and (4,3).
Step 4 IC L(C 3 ) = 0.8013 > 0.8, SC L(C 3 ) = 0.9107 > 0.9, then C 3 ∈
Expert e19 should increase his assessments for positions (2, 1 )
S − HH;
and (2, 4 ), and decrease his assessment for position (4, 2 ).
Step 4 IC L(C 4 ) = 0.7825 < 0.8, SC L(C 4 ) = 0.8860 < 0.9, then C 4 ∈
Expert e20 should increase his assessments for positions (2, 1 ),
S − LL.
Step 5 Hybrid strategies are used to help experts modify their
(2, 4 ), and (4,3), and decrease his assessment for position
preferences. Because C 1 , C 3 ∈ S−H H , the experts in sub-
( 4, 2 ).
groups C 1 and C 3 do not need to adjust their preferences. We should note that the adjustment suggestions may require
the experts to make many changes with respect to some pairs
1) Because C2
∈ S−LH, the feedback strategy for C2 is implemented
of alternatives. In these situations, the experts may completely
as follows:
refuse to follow the suggestions, and instead, make the modifi-
γ S−LH = 0.7648 {ca1 = 0.7318(c p12 = 0.9067, c p13 = 0.6054, cations that they like. This involves another important topic in
c p14 = 0.6832 ); LSGDM problems, that is, non-cooperative behavior management.
ca2 = 0.7920(c p21 = 0.9067, c p23 = 0.7201, c p24 = 0.7491 ); There are some studies that focused on detecting and managing
ca3 = 0.7500(c p31 = 0.6054, c p32 = 0.7201, c p34 = 0.9243 ); non-cooperative behaviors (Palomares et al., 2014; Xu, Du et al.,
ca4 = 0.7855(c p41 = 0.6832, c p42 = 0.7491, c p43 = 0.9243 )}. 2015; Xu, Du, Chen & Cai, 2019). Since this is not the emphasis
of our study, we assume that all experts agree to modify their
i) According to Eq. (10), the alternatives that should be adjusted preferences.
are x1 , x3 . According to the previous recommendations, experts provide
ii) According to Eq. (11), the positions that should be adjusted are the modified preferences as follows:
⎧ ⎫ ⎧ ⎫
(1,3), (1,4), (3,1), and (3,2). ⎪ 0.4⎪ ⎪ 0.2 ⎪
⎨0.5 0.4 0.2
⎬ ⎨0.5 0.4 0.4
⎬
0.6 0.5 0.3 0.4 0.6 0.5 0.5 0.25
Generation of advice: P7 = 8
,P = ,
All the experts in sub-group C 2 should increase their as- ⎩0.8
⎪ 0.7 0.5 0.6⎪
⎭ ⎩0.6
⎪ 0.5 0.5 0.4 ⎪
⎭
0.6 0.6 0.4 0.5 0.8 0.75 0.6 0.5
sessments for positions (1,3) and (1,4), and decrease their ⎧ ⎫ ⎧ ⎫
assessments for positions (3,2) and (3,2). ⎪
⎨0.5 0.6 0.2 0.4⎪
⎬ ⎪
⎨0.5 0.6 0.6 0.4⎪
⎬
0.4 0.5 0.3 0.2 0.4 0.5 0.4 0.2
P9 = 16
,P =
2) Because C 4 ∈ S − LL, the feedback strategy for C 4 is imple- ⎪0.8
⎩ 0.7 0.5 0.6⎪
⎭ ⎩0.4
⎪ 0.6 0.5 0.2⎪
⎭
mented as follows: 0.6 0.8 0.4 0.5 0.6 0.8 0.8 0.5
⎧ ⎫ ⎧ ⎫
i) According to Eq. (13), the experts who should make modifi- ⎪
⎨0.5 0.8 0.7 0.6⎪
⎬ ⎪
⎨0.5 0.7 0.8 0.2⎪
⎬
0.2 0.5 0.4 0.3 0.3 0.5 0.6 0.4
cations are EX P S−LL = {e16 , e18 , e17 , e19 , e20 }. P 17 = , P 18 = ,
ii) According to Eq. (14), the alternatives that should be ⎩0.3
⎪ 0.6 0.5 0.4⎪
⎭ ⎩0.2
⎪ 0.4 0.5 0.2⎪
⎭
0.4 0.7 0.6 0.5 0.8 0.6 0.8 0.5
adjusted are listed in Table 4. ⎧ ⎫ ⎧ ⎫
iii) According to Eq. (15), the positions that should be adjusted ⎪
⎨0.5 0.8 0.7 0.3 ⎪
⎬ ⎪
⎨0.5 0.8 0.8 0.6⎪
⎬
are listed in Table 4. 0.2 0.5 0.4 0.3 0.2 0.5 0.3 0.2
P 19 = 20
,P =
⎩0.3
⎪ 0.6 0.5 0.15⎪
⎭ ⎩0.2
⎪ 0.7 0.5 0.3⎪
⎭
Generation of advice: 0.7 0.7 0.85 0.5 0.4 0.8 0.7 0.5
M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971 967
Table 5
Weights of clusters (second round).
C1 0.6733 6 0.3346
C2 0.5337 3 0.1455
C3 0.4238 6 0.2607
C4 0.6000 5 0.2592
Second round
Using the proposed FCM clustering algorithm, the clustering
results are unchanged.
Then, the weights of the clusters are calculated and presented
in Table 5.
Values of ICL for four subgroups are ICL(C1 ) = 0.8428,
ICL(C2 ) = 0.7706, ICL(C3 ) = 0.8200, ICL(C4 ) = 0.8097. We can
Fig. 7. Relationship between the number of clusters and the silhouette coefficient.
obtain GCL = 0.8108 > 0.8. Thus, the group reaches an acceptable
consensus level. The global collective RCM is
⎧ ⎫ experts who form a sub-group on their own (minority opin-
⎪
⎨ 0.5 0.5450 0.5696 0.5542⎪
⎬
0.4550 0.5 0.5597 0.5649 ions) will have a larger weight than the experts who have team
G
P = . companions. There are some studies that focused on manag-
⎪0.4304 0.4403 0.5 0.4665⎪
⎩ ⎭ ing minority opinions in LSGDM problems (Xu, Du et al., 2015).
0.4458 0.4351 0.5335 0.5
Considering all these factors, K = 4 is an appropriate choice.
Then, using the normalizing rank aggregation method (Xu & Da, 3) The final alternative ranking is stable with the change of K.
2005), we can obtain the final ranking of these four alternatives However, there are some differences on the final group RCM.
as x1 > x2 > x4 > x3 . The neutralization effect also contributes to the fact that these
differences are not particularly significant since some experts
5. Discussions and comparative analyses increase the assessments and some experts decrease the assess-
ments.
In this section, we provide some discussions about setting
Rodríguez et al. (2018) and Labella, Liu, Rodríguez and Martínez
parameters for the proposed model and the comparisons with
(2019) also applied the idea of adaptive CRP to solve LSGDM prob-
related LSGDM methods.
lems. To demonstrate the generalizability and applicability of our
proposed model, we also provide the performance of Rodríguez
5.1. Analysis of K et al. (2018)’s and Labella et al. (2019)’s methods. Both of these
two studies set two thresholds regarding the level of consensus.
The clustering process plays a critical role in LSGDM. It can The first one is the level of consensus for advice generation, i.e.,
reduce the dimension of large-scale experts, based on which the ς = 0.7: if the level of global consensus is lower than 0.7, then
complexity and cost can be reduced (Ma, Zhu, Ponnambalam & all experts in sub-groups with their proximity degrees lower than
Zhang, 2019). Furthermore, it can identify the spokesperson repre- the average proximity degree are selected. The proximity degree
senting the subgroup through finding common opinion patterns in Rodríguez et al. (2018) is equal to the inter consensus level
such as sub-groups with similar opinions (Palomares, 2018). In the ICL in this paper. Another parameter is the consensus threshold,
FCM algorithm, the initial number of clusters K should be set in i.e., θ = 0.85. If ς ≤ GCL < θ , then the level of consensus is high
advance. To better reflect the impact of the number of clusters but not sufficient. In this case, the selection of sub-groups is the
on the decision results, we simulate the CRP and present the same as that in the above case. The difference from the above
results in Table 6. Based on Table 6, we can obtain the following case is that only those experts whose opinions differ most from
observations: the collective one are selected. If GCL ≥ θ , then the degree of
consensus is sufficiently high.
1) The value of initial GCL decreases as the number of clusters
Since Rodríguez et al. (2018) took the average proximity degree
increases. When the number of clusters is not greater than 3,
as the sub-group selection criterion, nearly half of the sub-groups
the initial value of GCL is larger than 0.8. Therefore, there is
were selected in each round regardless the degree of consensus. By
no need to provide modification suggestions to experts. How-
contrast, our method covers a wide range of clusters if the degree
ever, the silhouette coefficients are very low in these two situ-
of consensus is low, and a narrow range if the degree of consensus
ations. Fig. 7 presents the relationship between the silhouette
is high. From Table 6, we can summarize two advantages of our
coefficient and the number of clusters. As mentioned before,
method:
the silhouette coefficient is an indicator to measure the clus-
√
tering effect. A higher value of silhouette coefficient reflects a First, when the number of clusters is 6 or 7 with a low initial
better clustering effect. The silhouette coefficient is low when GCL, our method has a higher iterative efficiency since more ex-
the number of clusters is 2 or 3 because the partition is rough. perts are covered in the first iteration.
√
2) There is a rapid growth for the silhouette coefficient from 3 Second, when the number of clusters is 4 or 5, both of our
clusters to 4 clusters and small changes between 4 and 6. We method and Rodríguez et al. (2018)’s method need only one it-
should note that there are clusters that contain only one ex- eration. However, our method requires less modifications since
pert when the number of clusters is greater than 4. In such a less pairs of alternatives are covered according to the adaptive
case, the value of A(P k ) for ek is 1 and S(P k ) = 1. Furthermore, feedback mechanism presented in Section 3.4. Therefore, our
we set the value of SCI for these clusters as 1. Therefore, those method has a better performance in avoiding over-adjustment
968 M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971
Table 6
Decision results associated with the number of clusters.
K Initial clustering results Initial NI(1) NP(1) NI(2) NP(2) Final Final global group RCM Alternative ranking
GCL GCL
⎧ ⎫
⎪0.50 0 0 0.5611 0.5575 0.5242⎪
⎨ ⎬
0.4389 0.50 0 0 0.5391 0.5237
2 C 1 = {e1 , e2 , e3 , e4 , e5 , e6 , e11 , e12 , 0.8660 0 0 0 0 0.8660 x1 > x2 > x4 > x3
e13 , e14 , e17 , e18 , e19 , e20 },
⎪
⎩0.4425 0.4609 0.50 0 0 0.4546⎪
⎭
0.4758 0.4763 0.5454 0.50 0 0
C 2 = {e7 , e8 , e9 , e10 , e15 , e16 }
⎧ ⎫
⎪0.50 0 0 0.5437 0.5459 0.5271 ⎪
⎨ ⎬
0.4563 0.50 0 0 0.5435 0.5447
3 C 1 = {e1 , e2 , e3 , e4 , e5 , e6 , 0.8137 0 0 0 0 0.8137 x1 > x2 > x4 > x3
e17 , e18 , e19 , e20 },
⎪
⎩ 0.4541 0.4565 0.50 0 0 0.4673 ⎪
⎭
0.4729 0.4553 0.5327 0.50 0 0
C 2 = {e7 , e8 , e9 , e16 },
C 3 = {e10 , e11 , e12 , e13 , e14 , e15 }
⎧ ⎫
⎪0.50 0 0 0.5450 0.5696 0.5542⎪
⎨ ⎬
0.4550 0.50 0 0 0.5597 0.5649
4 C1 = {e1 , e2 , e3 , e4 , e5 , e6 }, 0.7888 1 23 1 32 0.8108 x1 > x2 > x4 > x3
C 2 = {e7 , e8 , e9 },
⎪
⎩0.4304 0.4403 0.50 0 0 0.4665⎪
⎭
0.4458 0.4351 0.5335 0.50 0 0
C = {e10 , e11 , e12 , e13 , e14 , e15 },
3
Note. NI(1): number of iterations by using our model; NP(1): number of positions in the upper triangle of RCM need to be modified by using our model; NI(2): number of
iterations by using Rodríguez et al. (2018)’s model; NP(2): number of positions in the upper triangle of RCM need to be modified by using Rodríguez et al. (2018)’s model.
Table 7
Decision results associated with different consensus thresholds.
Consensus NI(1) NP (NE)(1) NI(2) NP(NE)(2) Final global group RCM Alternative
threshold ranking
(GCI)
⎧ ⎫
⎪0.50 0 0 0.5600 0.5613 0.5334 ⎪
⎨ ⎬
0.4400 0.50 0 0 0.5465 0.5343
0.75 0 0 0 0 x1 > x2 > x4 >
⎪
⎩0.4387 0.4535 0.50 0 0 0.4577 ⎪⎭ x3
0.4666 0.4657 0.5423 0.50 0 0
⎧ ⎫
⎪0.50 0 0 0.5450 0.5696 0.5542⎪
⎨ ⎬
0.4550 0.50 0 0 0.5597 0.5649
0.8 1 23(8) 1 32(8) x1 > x2 > x4 >
⎪
⎩0.4304 0.4403 0.50 0 0 0.4665⎪ ⎭ x3
0.4458 0.4351 0.5335 0.50 0 0
⎧ ⎫
⎪0.50 0 0 0.5715 0.5905 0.5333 ⎪
⎨ ⎬
0.4285 0.50 0 0 0.5247 0.5581
0.85 2 59(20)+18(8) 4 32(8)+34(9) x1 > x2 > x4 >
+27(10)+26(10)
⎪
⎩0.4095 0.4753 0.50 0 0 0.4831 ⎪⎭ x3
0.4667 0.4419 0.5169 0.50 0 0
⎧ ⎫
⎪0.50 0 0 0.5491 0.5543 0.5171 ⎪
⎨ ⎬
0.4509 0.50 0 0 0.5346 0.5501
0.9 3 7 x1 > x2 > x4 >
59(20)+54(20)+24(9) 32(8)+34(9)+27(10)
⎪
⎩ 0.4457 0.4654 0.50 0 0 0.4425⎪ ⎭ x3
0.4829 0.4499 0.5575 0.50 0 0
+26(10)+26(9)
+22(8)+23(8)
Note. NE: number of experts that need to make modifications. NI(1): number of iterations by using our model; NP(NE)(1): number of positions (experts) in the upper triangle
of RCM need to be modified by using our model; NI(2): number of iterations by using Rodríguez et al. (2018)’s model; NP(NE)(2): number of positions (experts) in the upper
triangle of RCM need to be modified by using Rodríguez et al. (2018)’s model.
when the degree of consensus is close to the threshold. In this 5.2. Analysis of the consensus threshold
sense, the supervision cost of our method is reduced.
In this section, we analyze the effect of different consensus
Overall, our method can adapt to fine or rough partitioning thresholds on the decision results. Table 7 presents the results
scenarios. Furthermore, if there are one or more minority opinion including the number of iterations, number of adjusted positions,
subgroups, the CRP can also reach an expected consensus level final group RCM and alternative ranking. Based on Table 7, some
within one iteration. Hence, this method is robust to the partition observations can be obtained as:
distribution.
M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971 969
1) The number of iterations increases as the value of consensus consensus in our illustrative example is high but not sufficient. Us-
threshold increases, and the final alternative ranking is stable. ing the method in Rodríguez et al. (2018), two sub-groups (C 2 and
2) The closer the value of GCL to the consensus threshold is, the C 4 ) and all the four alternatives are selected. For sub-group C 2 , the
smaller the quantity of positions and experts involved in the pairs of alternatives (x1 , x3 ), (x1 , x4 ), (x2 , x3 ), and (x2 , x4 ) are se-
feedback mechanism are. For instance, when GCI = 0.85, a total lected; for sub-group C 4 , the pairs of alternatives (x1 , x2 ), (x1 , x3 ),
of 59 positions (in the upper triangle of the RCM) for all 20 ex- (x1 , x4 ), (x2 , x4 ), and (x3 , x4 ) are selected. If we focus on the cost
perts need to be modified in the first iteration round. In the of supervision, the moderator needs to provide 3 × 4 + 4 × 5 = 32
second iteration round, 8 experts should make modifications suggestions for these seven experts. Our proposed model also
in 18 positions altogether. Furthermore, in the same iteration needs to provide 32 suggestions. However, these 32 suggestions
round, if the consensus thresholds are different, the numbers of cover all 4 sub-groups and 20 experts. Because C 1 , C 2 , C 3 ∈ S−LH,
positions and experts involved in the feedback mechanism are then all experts in these three sub-groups are required to change
also different. For instance, when GCI = 0.8, 8 experts should following a unified direction. After the first iteration, our method
make modifications in 23 positions in the first iteration round. reaches a degree of consensus of 0.8412 and Rodríguez et al.
However, when GCI = 0.85, all 20 experts should make modi- (2018)’s method reaches 0.8125.
fications in 59 positions in the first iteration round. With the Except these two adaptive CRPs, there are several other CRPs.
increase of GCI, some sub-groups belong to S-HH turn into S- One widely used CRP is to choose subgroups whose degree of con-
LH. These two perspectives can elaborate the word “adaptive” sensus is lower than a predefined threshold in each iterative round
in the title. and requires them to make modifications (Zhang, Dong, Chiclana
3) A higher consensus threshold GCI needs a larger quantity of & Yu, 2019) (CRP1). The other is to choose the furthest subgroup
modifications, which distorts the original opinions. For instance, in each iterative round (Xu, Du et al., 2015) (CRP2). When we use
when GCI = 0.9, all experts are identified to make modifications these four CRPs to solve the problem, the iterative processes are
in two rounds. This explains why an LSGDM problem should presented in Fig. 8. As we can see from Fig. 8, our method has a
not set up a high consensus threshold. The consensus thresh- high efficiency. CRP1 has three iterations. However, compared with
old should be set according to actual situations. the adaptive CRP, CRP1 has the problem of over-adjustment. After
the second iteration, CRP1 reaches a degree of consensus that is
From Table 7, we can also find the two advantages summarized close to 0.85. After the third iteration, its degree of consensus is
in Section 5.1 compared with Rodríguez et al. (2018)’s method. the highest because many modifications have been made. CRP2 is
One is the higher iterative efficiency. When GCI is high, such as the slowest since only the farthest subgroups are required to make
GCI =0.9, Rodríguez et al. (2018)’s method needs more iterations. modifications in each round.
The other is avoiding over-adjustment. As we can see from the Based on all above discussions, we can summarize the advan-
fifth column of Table 7, in each iterative round, nearly half of tages of the proposed model:
all experts are involved. One big difference between Rodríguez
et al. (2018)’s method and our method is that when the degree
of consensus is far from the threshold, then a large number
of modifications are required to fasten the iteration process. In
the initial iterations, all experts will participate in the feedback 1) It has a high iterative efficiency, especially in the initial itera-
mechanism. In the last iterations, only a small number of experts tions. The adaptive mechanism involves a large scope when the
and pairs of alternatives are involved. If we use Rodríguez et al. degree of consensus GCL is far from the threshold GCI.
(2018)’s method, one problem is that some subgroups may need 2) It can avoid over-adjustment when the degree of consensus GCL
to make adjustments of similar quantities in several successive is near to the threshold GCI because of the introduction of S−LH
iterations, which may discourage their enthusiasm to participate and S−HL.
in the feedback mechanism. Additionally, our method has a better These two advantages make the proposed model have a good
performance with the change of GCI and is relatively insensitive to performance with the change of K and GCI (see Tables 6 and
the selection of GCI. 7).
3) The introduction of S−LH can reduce supervision cost since
5.3. Comparison with different CRPs only a unified suggestion is required for the low–high sub-
groups.
Firstly, we use Rodríguez et al. (2018)’s method to solve the 4) The clusters in this model are allowed to change. In the CRP,
problem. Because ς ≤ GCL = 0.7888 < θ , then the initial degree of experts make modifications on their preferences. Therefore, the
970 M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971
clusters may change. The adoption of fixed clusters (Liu, Xu, moderator can be reduced. For the experts belonging to S-LH,
Montes, Ding & Herrera, 2019; Xu, Du et al., 2015) is a special a unified suggestion for the sub-group instead of specific sug-
case in the proposed model. gestion for each expert is enough. For the experts belonging to
5) The weight determination combines two indices, namely, the S-HL, the task of the moderator is to make these experts’ opin-
majority principle and cohesion of subgroups. The single index- ions more cohesive. In this situation, only a small number of
based method based on the majority principle (Xu, Du et al., changes are required since these experts have been classified
2015) or the degree of consistency of RCMs has some limita- into the same cluster.
tions. For instance, if we only use the majority principle, the
In this study, the preference structures that the experts used
subgroups with the same number of experts but possessing dif-
to express their opinions were assumed to be RCMs. In the future,
ferent inner characteristics will have the same weights. If we
it is necessary to investigate the LSGDM method in the case that
use only the degree of consistency, the importance of experts
the representation models are qualitative. In addition, this work
in a small subgroup with a high degree of consistency may be
mainly focused on the cost of supervision, i.e., the number of
several times higher than the importance of experts in large
suggestions provided by the moderator. The adjustment cost can
subgroups with a low degree of consistency. Therefore, the pro-
also be taken into account under the adaptive LSGDM context in
posed method is more applicable.
the future. The optimization of the time complexity of the LSGDM
There are many other consensus models for LSGDM problems consensus model is another interesting topic.
(Palomares et al., 2014; Wu & Xu, 2018; Xu, Du et al., 2015; Xu,
Zhong, Chen & Zhou, 2015; Zhang et al., 2018). Different models Acknowledgments
have their own purposes. Therefore, it is not fair to compare these
different consensus models. There is not a model that can be The work was supported by the National Natural Science Foun-
regarded as the best. dation of China (71571156 and 71971145), and the 2019 Sichuan
The proposed model also has limitations. Several parameters Planning Project of Social Science (No. SC18A007).
used in this study, including the consensus thresholds GCL and SCL,
the control parameter δ in Eq. (16), and the number of clusters K, Appendix
should be set in advance. For GCL and SCL, some discussions are
Twenty initial RCMs:
given in Sections 3.3 and 5.2. These two thresholds can be deter- ⎧ ⎫ ⎧ ⎫
mined by the moderator according to the specific decision-making ⎪
⎨0.5 0.7 0.9 0.8⎪
⎬ 0.7 0.85 ⎪
⎨ 0.5 0.8⎪
⎬
problem. The value of δ controls the degree of preference between 1 0.3 0.5 0.8 0.7 2 0.3 0.5 0.7 0.6
P = ,P = ,
⎩0.1 0.2 0.5 0.4⎪
⎪ ⎩0.15 0.3 0.5
⎪ 0.3⎪
the global collective point and the center point of a cluster. It can
⎭ ⎭
be determined by the degrees of deviation and diversity of the 0.2 0.3 0.6 0.5 0.2 0.4 0.7 0.5
group. The number of clusters, K, is also a parameter that should ⎧ ⎫ ⎧ ⎫
be considered. The number of clusters must be first defined for ⎪
⎨0.5 0.7 0.8 0.6⎪
⎬ 0.6 0.65 0.75⎪
⎬ ⎪
⎨ 0.5
3 0.3 0.5 0.8 0.6 4 0.4 0.5 0.6 0.8
many automatic clustering algorithms. In Section 5.1, we discuss P = ,P = ,
the appropriate number of clusters with the help of silhouette ⎩0.2 0.2 0.5 0.4⎪
⎪ ⎭ ⎩0.35 0.4 0.5 0.55⎪
⎪ ⎭
0.4 0.4 0.6 0.5 0.25 0.2 0.45 0.5
coefficient and minority opinions. There are also many studies that ⎧ ⎫ ⎧ ⎫
focused on this topic and some approaches have been developed, ⎪
⎨ 0.5
0.7 0.8 0.75⎪
⎬ ⎪
⎨ 0.5 0.7 0.85 0.8 ⎪
⎬
such as the validity indices (Wu & Yang, 2005) and the robust 5 0.3 0.5 0.7 0.6 6 0.3 0.5 0.7 0.55
P = ,P = ,
and adaptive clustering analysis method (Mok, Huang, Kwok &
⎩ 0.2 0.3
⎪ 0.5 0.4 ⎪
⎭ ⎩0.15 0.3 0.5 0.4 ⎪
⎪ ⎭
Au, 2012). The different methods have their own advantages. From 0.25 0.4 0.6 0.5 0.2 0.45 0.6 0.5
another perspective, these parameters provide great freedom for ⎧ ⎫ ⎧ ⎫
the moderator to control the decision-making process.
⎪
⎨0.5 0.4 0.1 0.3⎪
⎬ 0.4 ⎪
⎨ 0.5 0.3 0.05⎪
⎬
7 0.6 0.5 0.2 0.4 8 0.6 0.5 0.4 0.25
P = ,P = ,
6. Conclusions ⎩0.9 0.8 0.5 0.6⎪
⎪ ⎭ ⎩ 0.7 0.6
⎪ 0.5 0.4 ⎪
⎭
0.7 0.6 0.4 0.5 0.95 0.75 0.6 0.5
⎧ ⎫ ⎧ ⎫
This study developed an adaptive consensus model with hy- ⎪
⎨0.5 0.6 0.1 0.3⎪
⎬ ⎪
⎨0.5 0.3 0.1 0.4 ⎪
⎬
brid strategies within the LSGDM context. A key characteristic 0.4 0.5 0.2 0.2 0.7 0.5 0.3 0.6
P9 = , P 10 = ,
of this model is the design of a hybrid feedback mechanism to ⎩0.9 0.8 0.5 0.6⎪
⎪ ⎭ ⎩0.9 0.7 0.5 0.65⎪
⎪ ⎭
improve the consensus performance in LSGDM problems. This 0.7 0.8 0.4 0.5 0.6 0.4 0.35 0.5
model is composed of four parts: classifying experts, determining
⎧ ⎫ ⎧ ⎫
⎪
⎨0.5 0.3 0.4 0.6⎪⎬ ⎪
⎨ 0.5 0.25 0.4 0.55⎪
⎬
the weights of sub-groups, computing degrees of consensus, and 0.7 0.5 0.6 0.9 0.75 0.5 0.7 0.8
improving levels of consensus. An illustrative example about the P 11 = , P 12 = ,
legislative amendment in UNCITRAL was implemented to illus- ⎩0.6 0.4 0.5 0.7⎪
⎪ ⎭ ⎩ 0.6 0.3 0.5 0.65⎪
⎪ ⎭
0.4 0.1 0.3 0.5 0.45 0.2 0.35 0.5
trate the method. The main contributions of this study can be ⎧ ⎫ ⎧ ⎫
highlighted as follows: ⎪
⎨ 0.5 0.3 0.45 0.6⎪
⎬ ⎪
⎨0.50.4 0.4 0.6 ⎪
⎬
0.7 0.5 0.55 0.9 0.6 0.5 0.7 0.95
1) A weight-determining method for the sub-groups was devel- P 13 = , P 14 = ,
. .45 0.5 0.7⎪
oped. This method combines two parameters: cohesion and size
⎪
⎩0 55 0
⎭ ⎩0.6 0.3
⎪ 0.5 0.6 ⎪
⎭
0.4 0.1 0.3 0.5 0.4 0.05 0.4 0.5
of a sub-group. ⎧ ⎫ ⎧ ⎫
2) The proposed model truly takes into account the impact of the ⎪
⎨ 0.5 0.2 0.3 0.45⎪
⎬ ⎪
⎨0.5 0.7 0.6 0.4⎪
⎬
clustering process in LSGDM problems. Because of the cluster- 0.8 0.5 0.6 0.7 0.3 0.5 0.4 0.1
P 15 = , P 16 = ,
ing algorithm, an LSGDM problem was divided into two levels: ⎩ 0.7 0.4
⎪ 0.5 0.6 ⎪
⎭ ⎩0.4
⎪ 0.6 0.5 0.2⎪
⎭
inter sub-group and intra sub-group. According to the different 0.55 0.3 0.4 0.5 0.6 0.9 0.8 0.5
⎧ ⎫ ⎧ ⎫
levels of degree of consensus in the two levels, four scenarios ⎪
⎨0.5 0.8 0.7 0.6⎪
⎬ ⎪
⎨0.5 0.6 0.9 0.2⎪
⎬
were distinguished. Then, hybrid strategies were generated to 0.2 0.5 0.4 0.2 0.4 0.5 0.7 0.3
P 17 = , P 18 = ,
⎩0.3 0.6 0.5 0.3⎪ ⎩0.1 0.4 0.5 0.1⎪
provide suggestions for experts to modify their opinions. Be- ⎪ ⎪
⎭ ⎭
cause of this hybrid mechanism, the cost of supervision of the 0.4 0.8 0.7 0.5 0.8 0.3 0.9 0.5
M. Tang, H. Liao and J. Xu et al. / European Journal of Operational Research 282 (2020) 957–971 971
⎧ ⎫ ⎧ ⎫
⎪
⎨ 0.5 0.95 0.7 0.3 ⎪
⎬ ⎪
⎨0.5 0.9 0.8 0.6⎪
⎬
Mok, P. Y., Huang, H. Q., Kwok, Y. L., & Au, J. S. (2012). A robust adaptive clustering
analysis method for automatic identification of clusters. Pattern Recognition, 45,
0.05 0.5 0.4 0.2 0.1 0.5 0.3 0.1
P 19 = 20
,P = 3017–3033.
0.3 0.6 0.5 0.15⎪
⎪
⎩ ⎭ ⎩0.2
⎪ 0.7 0.5 0.4⎪
⎭ Pal, N. R., & Bezdek, J. C. (1995). On cluster validity for the fuzzy c-mean model.
0.7 0.8 0.85 0.5 0.4 0.9 0.6 0.5 IEEE Transaction on Fuzzy Systems, 3, 370–379.
Palarea-Albaladejo, J., Martín-Fernández, J. A., & Soto, J. A. (2012). Dealing with dis-
tances and transformations for fuzzy c-means clustering of compositional data.
References
Journal of Classification, 29(2), 144–169.
Palomares, I. (2018). LGDM approaches and models: A literature review. Large group
Bezdek, J. C., Ehrlich, R., & Wull, F. (1984). FCM – The fuzzy c-means clustering decision making. SpringerBriefs in computer science. Cham: Springer.
algorithm. Computers & Geosciences, 10(2–3), 191–203. Palomares, I., Martinez, L., & Herrera, F. (2014). A consensus model to detect and
Bezdekm, J. C. (1976). A physical interpretation of fuzzy ISODATA. IEEE Transactions manage non-cooperative behaviors in large-scale group decision making. IEEE
on Systems, Man and Cybernetics, 6, 387–390. Transaction on Fuzzy Systems, 22(3), 516–530.
Capuano, N., Chiclana, F., Fujita, H., & Herrera-Viedma, E. (2018). Fuzzy group deci- Pang, J. F., Liang, J. Y., & Song, P. (2017). An adaptive consensus method for multi-at-
sion making with incomplete information guided by social influence. IEEE Trans- tribute group decision making under uncertain linguistic environment. Applied
action on Fuzzy Systems, 26(3), 1704–1718. Soft Computing, 58, 339–353.
Chen, S. M., Lee, L. W., Yang, S. W., & Sheu, T. W. (2012). Adaptive consensus support Pérez, I. J., Cabrerizo, F. J., Alonso, S., & Herrera-Viedma, E. (2014). A new consen-
model for group decision making systems. Expert Systems with Applications, 39, sus model for group decision making problems with non-homogeneous experts.
12580–12588. IEEE Transactions on Systems, Man and Cybernetics: Systems, 44(4), 494–498.
Cheng, D., Zhou, Z. L., Cheng, F. X., Zhou, Y. F., & Xie, Y. J. (2018). Modeling the min- Pérez, I. J., Cabrerizo, F. J., & Herrera-Viedma, E. (2010). A mobile decision support
imum cost consensus problem in an asymmetric costs context. European Journal system for dynamic group decision-making problems. IEEE Transactions on Sys-
of Operational Research, 270, 1122–1137. tems, Man, and Cybernetics – Part A: Systems and Humans, 40, 1244–1256.
Dong, Q. X., & Cooper, O. (2016). A peer-to-peer dynamic adaptive consensus reach- Pérez, I. J., Cabrerizo, F. J., & Herrera-Viedma, E. (2011). Group decision making prob-
ing model for the group AHP decision making. European Journal of Operational lems in a linguistic and dynamic context. Expert Systems with Applications, 38(3),
Research, 250, 521–530. 1675–1688.
Efremov, R., Insua, D. R., & Lotov, A. (2009). A framework for participatory decision Rodríguez, A., & Laio, A. (2014). Clustering by fast search and find of density peaks.
support using Pareto frontier visualization, goal identification and arbitration. Science, 344(6191), 1492–1496.
European Journal of Operational Research, 199(2), 459–467. Rodríguez, R. M., Labella, Á., Tré, G. D., & Martínez, L. (2018). A large scale consen-
Gonzalez, T. F. (1984). Clustering to minimize the maximum inter-cluster distance. sus reaching process managing group hesitation. Knowledge-Based Systems, 159,
Theoretical Computer Science, 38(2–3), 293–306. 86–97.
Gupta, M. (2018). Consensus building process in group decision making—An adap- Rousseeuw, P. J. (1987). Silhouettes: A graphical aid to the interpretation and vali-
tive procedure based on group dynamics. IEEE Transaction on Fuzzy Systems, dation of cluster analysis. Journal of Computational and Applied Mathematics, 20,
26(4), 1923–1933. 53–65.
Herrera-Viedma, E., Cabrerizo, F. J., Kacprzyk, J., & Pedrycz, W. (2014). A review of Switalski, Z. (1999). Rationality of fuzzy reciprocal preference relations. Fuzzy Sets
soft consensus models in a fuzzy environment. Information Fusion, 17(1), 4–13. and Systems, 107, 187–190.
Herrera-Viedma, E., Herrera, F., Chiclana, F., & Luque, M. (2004). Some issues on con- Tian, J. F., Zhang, Z. M., & Ha, M. H. (2018). An additive consistency and consensus-
sistency of fuzzy preference relations. European Journal of Operational Research, based approach for uncertain group decision making with linguistic preference
154, 98–109. relations. IEEE Transaction on Fuzzy Systems. doi:10.1109/TFUZZ.2018.2865132.
Herrera-Viedma, E., Martínez, L., & Mata, F. (2005). A consensus support system Wu, K. L., & Yang, M. S. (2005). A cluster validity index for fuzzy clustering. Pattern
model for group decision-making problems with multigranular linguistic pref- Recognition Letters, 26(9), 1275–1291.
erence relations. IEEE Transaction on Fuzzy Systems, 13(5), 644–658. Wu, X. L., & Liao, H. C. (2019). A consensus-based probabilistic linguistic gained and
Kacprzyk, J., & Fedrizzi, M. (1986). Soft consensus measures for monitoring real lost dominance score method. European Journal of Operational Research, 272(3),
consensus reaching processes under fuzzy preferences. Control and Cybernetics, 1017–1027.
15(3–4), 309–323. Wu, Z. B., & Xu, J. P. (2016). Managing consistency and consensus in group decision
Kacprzyk, J., & Fedrizzi, M. (1988). A ‘soft’ measure of consensus in the setting making with hesitant fuzzy linguistic preference relations. Omega, 65(3), 28–
of partial (fuzzy) preferences. European Journal of Operational Research, 34(3), 40.
316–325. Wu, Z. B., & Xu, J. P. (2018). A consensus model for large-scale group decision mak-
Kacprzyk, J., & Fedrizzi, M. (1989). A ‘human-consistent’ degree of consensus based ing with hesitant fuzzy information and changeable clusters. Information Fusion,
on fuzzy logic with linguistic quantifiers. Mathematical Social Sciences, 18(3), 41, 217–231.
275–290. Xu, X. H., Du, Z. J., & Chen, X. H. (2015). Consensus model for multi-criteria large–
Kim, J. (2008). A model and case for supporting participatory public decision mak- group emergency decision making considering non-cooperative behaviors and
ing in e-democracy. Group Decision and Negotiation, 17(3), 179–192. minority opinions. Decision Support Systems, 79, 150–160.
Labella, Á., Liu, Y., Rodríguez, R. M., & Martínez, L. (2018). Analyzing the perfor- Xu, X. H., Du, Z. J., Chen, X. H., & Cai, C. G. (2019). Confidence consensus-based
mance of classical consensus models in large scale group decision making: A model for large-scale group decision making: A novel approach to managing
comparative study. Applied Soft Computing, 67, 677–690. non-cooperative behaviors. Information Sciences, 477, 410–427.
Labella, Á., Liu, Y., Rodríguez, R. M., & Martínez, L. (2019). An adaptive consensus Xu, X. H., Zhong, X. Y., Chen, X. H., & Zhou, Y. J. (2015). A dynamical consensus
reaching process dealing with comparative linguistic expressions in large-scale method based on exit-delegation mechanism for large group emergency deci-
group decision making. In Proceedings of the 11th conference of the European so- sion making. Knowledge-Based Systems, 86, 237–249.
ciety for fuzzy logic and technology. Atlantis Press. Xu, Z. S., & Da, Q. L. (2005). A least deviation method to obtain a priority vec-
Liang, Q., Liao, X. W., & Liu, J. P. (2017). A social ties-based approach for group de- tor of a fuzzy preference relation. European Journal of Operational Research, 164,
cision-making problems with incomplete additive preference relations. Knowl- 206–216.
edge-Based Systems, 119, 68–86. Zhang, H. J., Dong, Y. C., Chiclana, F., & Yu, S. (2019). Consensus efficiency in group
Liu, F., Zhang, W. G., & Zhang, L. H. (2014). Consistency analysis of triangular fuzzy decision making: A comprehensive comparative study and its optimal design.
reciprocal preference relations. European Journal of Operational Research, 235, European Journal of Operational Research, 275(2), 580–598.
718–726. Zhang, H. J., Dong, Y. C., & Herrera-Viedma, E. (2018). Consensus building for the
Liu, X., Xu, Y. J., Montes, R., Ding, R. X., & Herrera, F. (2019). Alternative rank- heterogeneous large-scale GDM with the individual concerns and satisfactions.
ing-based clustering and reliability index-based consensus reaching process for IEEE Transaction on Fuzzy Systems, 28(2), 884–898.
hesitant fuzzy large scale group decision making. IEEE Transaction on Fuzzy Sys- Zhang, S. T., Zhu, J. J., Liu, X. D., Chen, Y., & Ma, Z. Z. (2017). Adaptive consen-
tems, 27(1), 159–171. sus model with multiplicative linguistic preferences based on fuzzy information
Ma, Z. Z., Zhu, J. J., Ponnambalam, K., & Zhang, S. T. (2019). A clustering method granulation. Applied Soft Computing, 60, 30–47.
for large-scale group decision-making with multi-stage hesitant fuzzy linguistic Zhou, K. L., Fu, C., & Yang, S. L. (2014). Fuzziness parameter selection in fuzzy
terms. Information Fusion, 50, 231–250. c-means: The perspective of cluster validation. Science China Information Sci-
Mata, F., Martínez, L., & Herrera-Viedma, E. (2009). An adaptive consensus support ences, 57(11), 1–8.
model for group decision making problems in a multi-granular fuzzy linguistic
context. IEEE Transaction on Fuzzy Systems, 17(2), 279–290.