You are on page 1of 15

Journal of the Operational Research Society

ISSN: (Print) (Online) Journal homepage: https://www.tandfonline.com/loi/tjor20

Dynamic subgroup-quality-based consensus in


managing consistency, nearness, and evenness
quality indices for large-scale group decision
making under hesitant environment

Ming Tang, Huchang Liao, Xiaomei Mi, Xuanhua Xu & Francisco Herrera

To cite this article: Ming Tang, Huchang Liao, Xiaomei Mi, Xuanhua Xu & Francisco Herrera
(2021) Dynamic subgroup-quality-based consensus in managing consistency, nearness, and
evenness quality indices for large-scale group decision making under hesitant environment, Journal
of the Operational Research Society, 72:4, 865-878, DOI: 10.1080/01605682.2019.1708823

To link to this article: https://doi.org/10.1080/01605682.2019.1708823

Published online: 13 May 2020.

Submit your article to this journal

Article views: 94

View related articles

View Crossmark data

Citing articles: 2 View citing articles

Full Terms & Conditions of access and use can be found at


https://www.tandfonline.com/action/journalInformation?journalCode=tjor20
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY
2021, VOL. 72, NO. 4, 865–878
https://doi.org/10.1080/01605682.2019.1708823

ORIGINAL ARTICLE

Dynamic subgroup-quality-based consensus in managing consistency,


nearness, and evenness quality indices for large-scale group decision
making under hesitant environment
Ming Tanga , Huchang Liaoa,b , Xiaomei Mia , Xuanhua Xuc and Francisco Herrerab,d
a
Sichuan University, Chengdu, China; bUniversity of Granada, Granada, Spain; cCentral South University, Changsha, China; dKing
Abdulaziz University, Jeddah, Saudi Arabia

ABSTRACT ARTICLE HISTORY


With the increasing complexity of decision-making environment and the development of Received 16 April 2019
societal demands, large-scale group decision making (LSGDM) has become a hot topic in Accepted 17 December 2019
recent years. Due to the demand of the current decision-making environment, up to now,
KEYWORDS
studies regarding LSGDM methods focussing on different approaches have been published.
Large-scale group decision
However, there are few studies that focus on evaluating the results of LSGDM problems. making; adaptive consensus
This study proposes three subgroup quality indices, namely, consistency degree, nearness reaching process; subgroup
degree and evenness degree, to measure the effectiveness of LSGDM methods. These three quality indices; hesitant
quality indices not only can evaluate the effect of a subgroup on the global group, but also fuzzy set
can measure the effect of final decision results. Furthermore, the consistency measure is
used to design the weights of subgroups, based on which an adaptive dynamic consensus
reaching process is introduced. An illustrative example is given to verify the applicability and
effectiveness of our proposed model.

1. Introduction result, the uncertainty in decision making is becom-


ing increasingly inevitable. Fortunately, the proposal
Group decision making (GDM) can be viewed as a
of hesitant fuzzy set (HFS) (Torra, 2010) provides
situation where a group of experts express their
opinions regarding candidate alternatives, and then an efficient tool to deal with hesitation situations.
select the best one. Generally, a traditional GDM There is a necessity to use HFSs to express experts’
problem contains a small number of experts (Wu & uncertain information when they are hesitant among
Xu, 2018). However, with the increasing complexity several numerical values.
of social environments, there is a trend that a large Up to now, there are four main topics in the field
number of experts are involved in decision-making of LSGDM:
problems. Therefore, large-scale GDM (LSGDM)
becomes a much-talked-about research topic and 1) Sub-group clustering (Liu, Shen, Chen, Chen,
attracts a lot of attention from scientific scholars & Wang, 2014; Liu, Zhou, Ding, Palomares, &
(Ding, Wang, Shang, & Herrera, 2019; Ma, Zhu, Herrera, 2019; Wu & Xu, 2018);
Ponnambalam, & Zhang, 2019; Palomares, Martınez, 2) Consensus reaching process (CRP) in LSGDM
& Herrera, 2014; Song & Li, 2019). (Rodrıguez, Labella, De Tre, & Martınez, 2018;
An LSGDM problem has the following three Tang et al., 2019);
features (Ma et al., 2019): 1) The number of experts 3) LSGDM methods (Liu, Fan, & Zhang, 2016);
varies from dozens to hundreds; 2) Individual 4) Behaviour management (Li, Dong, & Herrera,
experts in the group can provide preferences in rela- 2019; Palomares et al., 2014; Xu, Du, &
tively different places and at relatively different Chen, 2015).
times; 3) Preferences provided by experts tend to be
uncertain due to the decision-making environment. Subgroup clustering is essential in LSGDM and
The third point is especially a challenging issue in generally has two effects (Ma et al., 2019): one is to
LSGDM. In LSGDM problems, large-scale experts reduce the size of the decision-making problem,
may have various backgrounds, expertise levels and based on which the cost and complexity can be
personal interests. Meanwhile, there are also hesitant reduced; the other is to find common opinion pat-
and uncertain nature in human cognition. As a terns such as subgroups with high similar judgment

CONTACT Huchang Liao liaohuchang@163.com Business School, Sichuan University, Chengdu 610064, China
ß The Operational Research Society 2020
866 M. TANG ET AL.

information, in which one member of the group can consensus model with a feedback mechanism that
be identified as a spokesperson. After the clustering can adjust the amount of advice, depending on the
process, we need to find a solution based on all sub- importance level of the expert. However, the weight
groups’ opinions. However, some subgroups may of each expert in their model remains static. In add-
think their opinions have not been fully considered ition, existing LSGDM models did not consider the
in the final solution and thus refuse to accept the CRP adapting to the weights of subgroups.
group decision result. Therefore, it is important to Therefore, it is an interesting topic to design an
apply the CRP in LSGDM to gain all subgroups’ adaptive consensus model with dynamic weights of
support. The CRP usually needs iterations to achieve experts (subgroups in LSGDM), which we called the
a minimum threshold. To reach this goal, all sub- adaptive dynamics CRP.
groups need to make their contributions. Owing to According to the above-mentioned facts, this
various perceptions, motivations and experience lev- study first proposes three subgroup quality indices
els, different subgroups may have different contribu- to evaluate the effect of LSGDM methods. These
tions and exert different influences on the global three subgroup indices are defined from three per-
group opinion. It seems of interest to measure the spectives: the first one is measuring the similarity
subgroup quality, i.e., a subgroup’s effect on the degree between priorities of subgroup’s alternatives
CRP, based on which, we can objectively determine with regard to that of the global collective group,
the weights of subgroups or adjust their weights which is called the consistency degree; the second
dynamically. one is the size of divorced values, which is called
LSGDM has been researched from different per- the nearness degree; the last one is the uniform dis-
spectives. However, studies from other aspects have tribution degree of total differences, which is called
barely been studied, such as measuring and evaluat- the evenness degree. Based on the judgment infor-
ing the effect of LSGDM results, or how to evaluate mation of subgroups, we first measure the opinion
the quality of a subgroup’s opinion in the final col- quality of each subgroup on the global group opin-
lective opinion. Currently, many approaches aiming ion, and then introduce a method to evaluate effects
at solving LSGDM problems have been developed. of decision results of LSGDM problems. Next, a
When we use different approaches to solve the same subgroup quality index named the consistency
problem, we may obtain different final rankings of degree is used to design the adaptive CRP in
alternatives, in other words, inconsistent decision LSGDM. In each iterative round, the weights of sub-
results are derived with respect to different LSGDM groups are adjusted dynamically. To reduce the
methods (Wu & Xu, 2018). The results can also be supervision cost, an adaptive mechanism with three
influenced by the clustering methods. Therefore, it strategies adapting to the weights of subgroups is
seems of interest to construct a quality index system put forward. The contributions of this study are
to evaluate the effect of LSGDM results, to know highlighted as follows:
which LSGDM method would be better from
an aspect. 1) We propose three subgroup quality indices, i.e.,
As aforementioned, we can design subgroups’ consistency degree, nearness degree and even-
dynamic weights in the CRP that are suitable for ness degree, to evaluate the influence of each
practical decision-making problems. The dynamic subgroup and the effect of LSGDM decision
consensus provides a tool to deal with dynamic results. It will be helpful to know the quality of
decision-making problems by allowing some param- decision results and select an LSGDM method
eters to be changed, as determined by external or with good performance.
subjective factors (Herrera-Viedma, Cabrerizo, 2) Based on the consistency degree, we introduce
Kacprzyk, & Pedrycz, 2014). In the decision-making a dynamic subgroup-quality-based CRP to
process, it seems of interest to analyse the environ- solve LSGDM problems. The weight of each
ment characteristics and conditions of the handled subgroup is determined by two factors: the
problem. In current situations, decision-making size and the consistency degree of the opin-
problems occur in a dynamic environment (Perez ions. Then, the weights of subgroups are
et al., 2018). So far, the dynamic consensus has been adjusted in the CRP dynamically. Three strat-
extended into different aspects. Among these egies adapting different weights of subgroups
aspects, the adaptive consensus (Perez, Cabrerizo, are introduced.
Alonso, & Herrera-Viedma, 2014) is an emerging
and efficient method. This type of consensus can The rest of this paper is organised as follows.
generate recommendations according to some Section 2 introduces the preliminaries used in this
parameters such as the level of consensus and the study. In Section 3, we put forward three subgroup
weights of experts. Perez et al. (2014) proposed a quality indices, i.e., consistency degree, nearness
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY 867

degree and evenness degree, to evaluate the results When evaluating alternatives, the preference rela-
of LSGDM problems. Section 4 presents details of tion based on pairwise comparisons is a popular
the adaptive dynamic CRP. In Section 5, an illustra- tool for experts to express their opinions. The con-
tive example and some comparisons are provided. cept of hesitant comparison matrix (HCM) com-
Conclusions are drawn in Section 6. posed of hesitant pairwise comparisons was
introduced (Xia & Xu, 2013). An HCM on X is rep-
resented as a matrix H ¼ ðhij Þnn  X  X, where
2. Preliminaries
hij ¼ fclij jl ¼ 1, 2, :::, #hij g (#hij is the number of ele-
This section introduces some preliminaries used in the ments in hij ) is an HFE, indicating all possible pref-
rest of this study, including the description of LSGDM
erence values to which alternative xi is preferred to
problems and the hesitant judgment information. rðlÞ rð#h lþ1Þ
xj : hij satisfies: cij þ cji ij ¼ 1, hii ¼ f0:5g,
rðlÞ
2.1. Description of large-scale group decision #hij ¼ #hji , for all i, j ¼ 1, 2, :::, n, where cij is the
making problems lth largest element in hij :
One problem in the decision-making process
Current societal and technological demands lead a
with hesitant judgment information is that the car-
large number of participants to take part in the
dinality of HFEs may be different. We adopt the
decision-making process (Tang et al., 2019). As a
parameter g proposed by Xu and Zhang (2013) to
result, LSGDM has become a hot research topic in
add elements to hij ði < jÞ, and 1g to add elements
recent years. The main elements of an LSGDM
to hji ði > jÞ: Then, a normalised HCM (NHCM)
problem are shown as follows:
(Zhu & Xu, 2014) H  ¼ ðh  ij Þ
nn can be obtained,
 ij ¼ maxf#h  ij ji, j ¼ 1, 2, :::, ng,
(1) A discrete set of alternatives X ¼ fx1 , x2 , which satisfies: #h
:::, xn g ðn  2Þ, which represent possible where #h  ij is the number of elements in hij :
solutions to the LSGDM problem; The score function of an HFE h is defined as
(2) A set of experts E ¼ fe1 , e2 , :::, em g, who are Farhadinia (2013):
invited to evaluate alternatives. There are two
X#h X #h
main principles for the value of m: One is sðhÞ ¼ dðlÞcl dðlÞ (2)
m  20 (Wu & Xu, 2018); the other is m  11 l¼1 l¼1
(Xu, Du, et al., 2015). In this study, we use #h is a positive-valued monotonically
where fdðlÞgl¼1
the former principle.
increasing sequence of the index l: Based on
The CRP in GDM or LSGDM is usually an itera- Equation (2), we can obtain a preference relation
tive process containing some discussion rounds. composed of score values. Then, the score value of
These discussion rounds need to be coordinated by each alternative can be derived as:
X
n 
the moderator, who plays a critical role in the CRP
sðxi Þ ¼ sðhij Þ n (3)
(Mata, Martınez, & Herrera-Viedma, 2009). The j¼1
tasks of the moderator mainly include (Palomares
et al., 2014): 1) calculating the consensus level of the To obtain the collective HCM for a group, the
group; 2) generating recommendations for experts weighted operator proposed by Xia and Xu (2013) is
to reach group consensus. used:
m
HFWAðh1 , h2 , :::, hm Þ ¼ 丣 ðwk hk Þ
k¼1
2.2. Hesitant judgment information X
m 
¼ [ wk crðsÞ
m
One feature of LSGDM problems is that the prefer- rðsÞ rðsÞ
c1 2h1 , c2 2h2 , :::, cm 2hm
rðsÞ
k¼1
ences provided by experts tend to be uncertain (Ma (4)
et al., 2019). Hence, we use the concept of HFS
T
(Torra, 2010) to depict uncertain and hesitant infor- where w ¼ ðw1 , w2 , :::, wm Þ is the weight vector of
P
mation. Let X be a reference set, an HFS on X can hk with wk 2 ½0, 1 and m k¼1 wk ¼ 1:
be represented in the form of
A ¼ fhx, hA ðxÞijx 2 Xg (1) 3. Subgroup quality indices for LSGDM:
where hA ðxÞ is a set of values in ½0, 1, denoting the
Consistency degree, nearness degree and
possible membership degrees of the element x 2 X
evenness degree
to the set A: hA ðxÞ is called the hesitant fuzzy elem- In this section, we introduce three quality sub-
ent (HFE) (Xia & Xu, 2013). group indices, i.e., consistency degree, nearness
868 M. TANG ET AL.

degree, and evenness degree, to evaluate the Through the normalisation process, we can obtain
results of LSGDM problems. NHCMs. To save space, four NHCMs are not illus-
trated. The collective HCM HG is obtained as:

8 9
>
> f0:5g f0:35, 0:4g f0:575, 0:625g >
f0:725g
>
>
> >
>
< f0:6, 0:65g f0:5g f0:625, 0:65, 0:675g f0:775, 0:85g =
HG ¼
>
> f0:375, 0:425g f0:325, 0:35, 0:375g f0:5g f0:65g > >
>
> >
>
: f0:275g f0:15, 0:225g f0:65g f0:5g ;

3.1. Consistency degree The score values of the alternatives corresponding


to each subgroup’s HCM and the collective HCM
An HCM composed of pairwise comparisons reflects
are shown in Table 1.
different preferences between alternatives. The con-
sistency degree of a subgroup’s HCM with respect
Based on Table 1, we can obtain the dominance
to the collective HCM can reflect the effect of the
class of each alternative regarding subgroup
subgroup on the global group opinion to some
extent. Inspired by Pang and Liang (2012), we use G1 :½x1   
s1 ¼ fx1 g, ½x2 s1 ¼ fx1 ,x2 g, ½x3 s1 ¼ fx1 ,x2 ,x3 g,
the dominance relation to measure the priorities ½x4 
s1 ¼ fx1 ,x2 ,x3 ,x4 g: Thus, the family set of domin-
between alternatives in this section. ance classes derived by the HCM HG1 is: X=R
s1 ¼
For a given HCM with nðn1Þ=2 pairwise com-
f½x1    
s1 ,½x2 s1 ,½x3 s1 ,½x3 s1 g ¼ ffx1 g,fx1 ,x2 g, fx1 ,x2 ,x3 g,
parisons among n alternatives, we say that alterna-
fx1 , x2 ,x3 ,x4 gg:
tive xi dominances xj if sðxi Þ  sðxj Þ, denoted as
Similarly, we can obtain the dominance class of
xi Rs xj : Then, the dominance class with regard to
each alternative derived by the collective HCM HG :
alternative xi is ½xi  s ¼ fxj 2 Xjsðxj Þ  sðxi Þg: Let ½x1   
sG ¼ fx1 , x2 g, ½x2 sG ¼ fx2 g, ½x3 sG ¼ fx1 , x2 , x3 g,
X=Rs ¼ f½x1 s , ½x2 s , :::, ½xn 
  
s g denote the family set ½x4 
sG ¼ fx1 , x2 , x3 , x4 g, and the family set of domin-
of dominance classes. The dominance classes in 
X=R ance classes derived by HG is X=R
sG ¼ f½x1 sG ,
s do not form a partition of X in general. We
use an example below to explain this concept. ½x2   
sG , ½x3 sG , ½x3 sG g ¼ ffx1 , x2 g, fx2 g, fx1 , x2 , x3 g, fx1 ,
x2 , x3 , x4 gg:
Example 1.1. Suppose that there is a solar power We use the first iterative round of the CRP to
company that needs to select a supplier from four explain the consistency index. Below we define the
possible alternatives: fx1 , x2 , x3 , x4 g: 20 experts are consistency degree of a subgroup HCM with regard
invited to evaluate these four suppliers. For the con- to the collective HCM.
venience of calculation, this example uses four sub- Definition 1. In an LSGDM problem, the consist-
groups ðG1 , G2 , G3 , G4 Þ with the same number of ency degree of an HCM HGh given by the subgroup
experts and the same weight. Four subgroups’ Gh with regard to the collective HCM HG is defined
HCMs are collected as follows: as

8 9 8 9
>
> f0:5g f0:6g f0:7, 0:8g f0:8g > > >
> f0:5g f0:2, 0:4g f0:6g f0:8g > >
>
> >
> >
> >
>
< f0:4g f0:5g f0:6g f0:6, 0:7g = < f0:6, 0:8g f0:5g f0:7g f0:8, 0:9g =
HG1 ¼ , HG2 ¼
> f0:2, 0:3g
> f0:4g f0:5g f0:6g > > >
> f0:4g f0:3g f0:5g f0:6g > >
>
> >
> >
> >
>
: f0:2g f0:3, 0:4g f0:4g f0:5g ; : f0:2g f0:1, 0:2g f0:4g f0:5g ;
8 9 8 9
>
> f0:5g f0:3g f 0:4, 0:5 g f0:6g > > >
> f0:5g f0:3g f0:6g f0:7g >
>
>
> >
> >
> >
>
< f0:7g f0:5g f0:6g f0:8, 0:9g = < f0:7g f0:5g f0:6, 0:7, 0:8g f0:9g =
HG3 ¼ , HG 4 ¼
> f0:5, 0:6g
> f0:4g f0:5g f0:7g > > >
> f0:4g f0:2, 0:3, 0:4g f0:5g f0:7g >
>
>
> >
> >
> >
>
: f0:4g f0:1, 0:2g f0:3g f0:5g ; : f0:3g f0:1g f0:3g f0:5g ;
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY 869

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Table 1. Score values of the alternatives corresponding to Pn 2
different HCMs. i¼1 ðsh ðxi ÞsG ðxi ÞÞ
Neah ¼ 1 (6)
x1 x2 x3 x4 n
G1 0.667 0.542 0.442 0.367
G2 0.558 0.700 0.450 0.317
G3 0.467 0.667 0.542 0.342 Obviously, 0 < Neah  1: The closer Neah is to 1,
G4 0.525 0.613 0.438 0.300 the higher the closeness of the subgroup Gh with
G 0.554 0.654 0.479 0.331
respect to the global group. As a result, this sub-
n j x  \ x j group’s opinion has a larger contribution to the cur-
1X ½ i sh ½ i sG
Conh ¼ (5) rent group opinion. In other words, this subgroup
n i¼1 j½xi 
s
j has a higher impact on the current iteration.
h

Example 1.3 (Continuation of Example 1.1). Based


where N is the number of subgroups. ½xi  
sh 2 X=Rsh on Definition 2, we can obtain the nearness degree of
denotes the dominance class of alternative xi each subgroup HCM with respect to the collect-
regarding the score values derived from the HCM HGh ive HCM.

sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ð0:6670:554Þ2 þ ð0:5420:654Þ2 þ ð0:4420:479Þ2 þ ð0:3670:331Þ2
Nea1 ¼ 1 ¼ 0:928
4

provided by the subgroup Gh : ½xi  


sG 2 X=RsG denotes
Similarly, Nea2 ¼ 0:972, Nea3 ¼ 0:946, Nea4 ¼
the dominance class of alternative xi regarding the 0:964: Based on the former consistency index,
collective values derived from the collective HCM HG : subgroups G2 and G4 have equal effects. If two
subgroups have the same consistency degree, then
Example 1.2 (Continuation of Example 1.1). The the subgroup with the higher nearness degree has
consistency degree of HG1 provided by the subgroup a higher impact on the group results, as the
G1 with respect to the collective HCM HG is: opinion of the subgroup is closer to the
group opinion. Therefore, the subgroup G2 is
4 j x  \ x j   more effective.
1X ½ i sh ½ i sG 1 1 1 3 4
Con1 ¼ ¼ þ þ þ
4 i¼1 j½xi 
s
j 4 1 2 3 4
h

¼ 0:875 3.3. Evenness degree


In this section, the concept of the evenness degree is
introduced based on the entropy measure. This
Similarly, the other three subgroups’ consistency
index measures the uniform distribution degree of
degrees are calculated: Con2 ¼ 1, Con3 ¼ 0:917,
total differences among all alternatives.
Con4 ¼ 1: Therefore, in this round, subgroups G2
and G4 have the largest effect on the group decision Definition 3. In an LSGDM problem, the evenness
under the consistency index. degree of the HCM HGh given by the subgroup Gh
Obviously, Conh 2 ð0, 1: The closer Conh is to 1, with regard to the global collective HCM HG is
the higher the consistency degree of the subgroup defined as:
Gh with respect to the global group. As a result, this
subgroup’s opinion has a larger contribution to the X
n
js ðx ÞsG ðxi Þj
Eveh ¼ K Pn h i
current group opinion. i¼1 i¼1 jsh ðxi Þ  sG ðxi Þj
(7)
jsh ðxi ÞsG ðxi Þj
ln Pn
i¼1 jsh ðxi Þ  sG ðxi Þj
3.2. Nearness degree
This section proposes another index to measure where K is the Boltzman’s constant (Nose, 1984),
subgroups’ effects in the LSGDM from the perspec- which equals ln1N such that 0 < Eveh  1:
tive of divorced values. Example 1.4 (Continuation of Example 1.1). Based
Definition 2. In an LSGDM problem, the nearness on Definition 3, we can obtain the evenness degree
degree of an HCM HGh given by the subgroup Gh of each subgroup’s HCM with regard to the global
with regard to the collective HCM HG is defined as collective HCM.
870 M. TANG ET AL.


1 0:113 0:113 0:112 0:112 original HCM HGh given by the subgroup Gh with
Eve1 ¼  ln þ ln
ln 40:298 0:298 0:298 0:298 ^ G:
respect to the final global collective HCM H

0:037 0:037 0:036 0:036
þ ln þ ln Definition 6. In an LSGDM problem, the total
0:298 0:298 0:298 0:298
evenness degree E of the results is defined as
¼ 0:902
X
N

Similarly, the other three subgroups’ evenness degrees E¼ vh Eve h (10)


h¼1
are: Eve2 ¼ 0:818, Eve3 ¼ 0:781, Eve4 ¼ 0:991:
Obviously, 0 < Eveh  1: The closer Eveh is to 1, where vh denotes the weight of the subgroup Gh
the higher the evenness degree of the distribution of and Eveh denotes the evenness degree of the original
difference between a subgroup’s score values and HCM HGh given by the subgroup Gh with respect to
the collective score values is. the final global collective HCM H ~ G:
Based on these three indices, we can evaluate the Obviously, all these three indices satisfy: C 2
effect of each subgroup in an LSGDM problem. In ð0, 1, P 2 ð0, 1 and E 2 ð0, 1: Using these three
an LSGDM problem, a subgroup with higher con- indices, we can evaluate the results of an LSGDM
sistency is closer to the global collective decision problem. First, for an LSGDM problem, if we use
results. If two subgroups have the same consistency different methods, then the method with the highest
degree, then the subgroup with the higher nearness total consistency degree indicates that all subgroups’
degree is closer to the global collective decision opinions are closer to the final decision results. If
results. If two subgroups have the same nearness two methods have the same consistency degree,
degree, then the subgroup with the higher evenness then the method with higher total nearness degree
degree is closer to the global collective deci- indicates that all subgroups’ opinions are closer to
sion results. the global decision results. If two methods have the
same nearness degree, then the method with the
highest evenness degree indicates that all subgroups’
3.4. Evaluating the results of an LSGDM problem opinions are closer to the final decision results.

In the following we introduce a method to evaluate


the results of an LSGDM problem, from which we 4. Subgroup quality-based dynamic CRP
can find out the degree of rationality and validity of
decision-making results. In this section, we propose a framework of the sub-
group quality-based CRP. This framework includes
Definition 4. In an LSGDM problem, the total con- four phases: 1) clustering process, 2) computing the
sistency degree C of the results is defined as consensus level, 3) consensus control and feed-
X
N back mechanism.
C¼ vh Con h (8)
h¼1
4.1. Clustering process
where vh denotes the weight of the subgroup Gh
The aim of a clustering algorithm is to classify data
and Con h denotes the consistency degree of the ori-
from a data set into finite and discrete sets accord-
ginal HCM HGh given by subgroup Gh with respect
ing to the similarity degrees between them.
^ G:
to the final global collective HCM H Elements in the same set should be more similar
The weight of a subgroup can be determined by than those in other sets. The clustering process is
its size (the number of experts). Most existing litera- essential for solving LSGDM problems.
ture uses this weight determination method. In this In 2014, Rodrıguez and Laio (2014) proposed a
study, we propose another method, which combines density-based clustering algorithm in Science,
the size and consistency degree to assign weight to named “fast search and find of density peaks
a subgroup. (FSFDP).” This algorithm introduced a new
Definition 5. In an LSGDM problem, the total method to select clustering centroids. The authors
nearness degree P of the results is defined as declared that a clustering centroid has the charac-
teristics of high local density and a larger relative
X
N
distance from other centroids. This algorithm is
P¼ vh Nea h (9)
h¼1
simple and efficient. In this study, the FSFDP algo-
rithm is applied to classify HCMs Hk , k ¼
where vh denotes the weight of the subgroup Gh 1, 2, :::, m: The procedure of the HCM-FSFDP algo-
and Nea h denotes the nearness degree of the rithm is presented as follows:
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY 871

determination process, we put the subgroup’s effect


Algorithm 1. (The HCM-FSFDP algorithm) in exponent in Equation (13).

Input: HCMs Hk ¼ ðhij, k Þnn , k ¼ 1, 2, :::, m and the


4.3. Consensus control
number of sub-groups NðN  2Þ:
Output: Subgroups G1 , G2 , :::, GN : In the process of GDM, the most critical issue is to
Step 1. Each HCM Hk is transformed into its nor- find a consensus result being supported by all
malised form, i.e., the NHCM. experts. Strictly speaking, consensus means a full
Step 2. Calculate the distances between all NHCMs. agreement among all experts regarding all alterna-
Here, we use the Euclidean distance (Xu & Xia, tives. However, this concept is hard to achieve in
2011) shown below to calculate the distances: real-world scenarios. Then, Saint and Lawson (1994)
0 0 111=2 softened this concept and described consensus as a
X
n #hxi
X
1 1 state of a mutual agreement among experts and all
dðH1 , H2 Þ ¼ @ @ jh ðxi Þ  h2 ðxi Þj AA
rðlÞ rðlÞ 2
n i¼1 #hxi l¼1 1 opinions have been conveyed to the satisfaction of
(11) them. A dynamic CRP can be defined as the process
of consensus changing with decision-making envir-
Step 3. Sort the distances dij in an ascending order onment and specific parameters (Perez et al., 2018).
and select a truncation distance dc : These environmental factors and parameters
Step 4. Compute the density qi of all nodes accord- include alternatives, experts, experts’ importance
ing to the function: and criteria.
X
qi ¼ vðdij dc Þ (12) Let fHG1 , HG2 , :::, HGN g be N HCMs given by N
j subgroups. The normalised forms of these N HCMs
are fH  G1 , H
 G2 , :::, H
 GN g: The collective matrix by
where vðxÞ is a 0-1 function such that: vðxÞ ¼ 1
aggregating N NHCMs is Hc ¼ ðhcij Þij : Then, the
when x < 0 and vðxÞ ¼ 0 when x  0: qi denotes
the number of nodes whose distance from Hi is consensus level of a subgroup is computed by
smaller than dc :  h Hc Þ
CLðGh Þ ¼ 1dðH (14)
Step 5. For each node Hi , find the node Hj that has
The global consensus level is obtained by
larger density than Hi and select the smallest dij ,
denoted as di : To select appropriate centroids, we 1X N
GCL ¼ CLðGh Þ (15)
set a threshold for the minimum value of qi and di : N h¼1
Step 6. Assign the remaining nodes to the same
cluster as its nearest neighbour of high density. The consensus degree on a position can be
Step 7. Output clusters G1 , G2 , :::, GN and end defined as
the algorithm. cphij ¼ 1jhhij hcij j (16)

The consensus degree on an alternative is


4.2. Dynamic weights based on quality indices Pn
j¼1, j6¼i cpij
After classifying the experts into several subgroups, cai ¼ (17)
n1
the next objective is to determine each subgroup’s
weight. In the first round of aggregating the HCMs A predefined consensus threshold d should be set
of subgroups into the collective HCM, we use the in advance. If GCL  d, the consensus process fin-
size of a subgroup as a principle to determine its ishes and the selection process is applied to obtain
weight. After that, we need to use the effect of each the final decision results. Otherwise, the feedback
subgroup to update its weight. strategy with another iterative discussion round
Let #Gh be the number of experts in subgroup Gh should be adopted to provide suggestions to experts.
and Conh be the consistency degree of Gh with Next, we describe the details of the adaptive strat-
respect to the global group. The size and consistency egy. All subgroups are classified into three levels: 1)
degree are aggregated as: Glow , low importance subgroups; 2)Gmed , medium
importance subgroups and 3) Ghigh , high import-
uðGh Þ ¼ ð1 þ #Gh ÞðConh minðConÞÞb (13)
ance subgroups. This classification is implemented
where minðConÞ is the minimum consistency degree by a fuzzy matching mechanism whose parameters
of all subgroups and b is a parameter that controls are determined according to specific decision-mak-
the impact of the consistency degree in the calculat- ing problems. We define three feedback strategies to
ing process. To highlight the importance of the sub- increase the consensus level. There are two phases:
group’s quality instead of size in the weight preference searching and advice generating.
872 M. TANG ET AL.

a) Preference searching for low-importance sub- the change of preferences. The moderator only pro-
groups. If a subgroup is of low importance, accord- vides modified suggestions to those experts whose
ing to our weight determination method, then this consensus level is lower than the average consensus
subgroup has a low consistency degree with the glo- degree of all experts.
bal group. In other words, the opinion of this sub-
group is far from the global group. It is reasonable i) For a subgroup Gh 2 Gmed , the experts whose
to require many modifications for these subgroups. consensus degree is lower than that of this
All experts in low-importance subgroups are subgroup as compared to the global group are
required to make modifications on all pairs of alter- identified by
natives that have low agreement levels. The search- EXP 2 ¼ fehr jCLhr < CLðGh ÞjGh 2 Gmed , ehr 2 Gh g
ing strategy for Glow acts as follows: (20)

i) For a subgroup Gh 2 Glow , the alternatives ii) The alternatives with a consensus degree lower
with a consensus degree lower than a thresh- than the threshold c2 , are identified by
old, c1 , are identified by ALT2 ¼ fxi jcai < c2 g (21)
ALT1 ¼ fxi jcai < c1 ÙGh 2 Glow g (18) where c2 can be calculated analogous to c1 :
The value of c1 can be set in advance or be
determined according to the specific consensus iii) The pairs of alternatives that need to be modi-
level in each discussion round. A high value of fied are identified by
c1 needs a large amount of changes while a POA2 ¼ fði, jÞjcai < c2 Ùxi 2 ALT2 g (23)
low value of c1 may increase iterative rounds.
iv) The set of preference values selected to make
Therefore, it is important to select an appro-
modifications is identified by
priate c1 : Here, we use the average degree of
consensus at the alternative level as c1 , such 2 ¼ fði, jÞ 2 POAjcai  ca
PREhr hr
 ig (24)
P
that c1 ¼ m i¼1 cai =m:
ii) The pairs of alternatives that need to be modi- Note that it is not necessary for all experts identi-
fied are identified by fied in Equation (20) to make changes. If an expert
does not have preference values that should be
POA1 ¼ fði, jÞjcai < c1 Ùxi 2 ALT1 g (19)
changed as identified by Equation (24), then this
expert does not need to make changes.
Generation of advices
Once the subgroups and pairs of alternatives that Generation of advices
need to make modifications are identified, the mod- The directions rules are recommended as follows:
erator provides direction rules to give suggestions. DR. 21: If sðhhr G
ij Þ < sðhij Þ, then the expert ehr in the
The directions rules are recommended as follows: subgroup Gh should increase the assessment
associated with the pair of alternatives ðxi , xj Þ:
DR. 11: If sðhhij Þ < sðhGij Þ, then all experts in the
subgroup Gh should increase the assessment G
DR. 22: If sðhhr
ij Þ > sðhij Þ, then the expert ehr in the
associated with the pair of alternatives ðxi , xj Þ:
subgroup Gh should decrease the assessment
associated with the pair of alternatives ðxi , xj Þ:
DR. 12: If sðhhij Þ > sðhGij Þ, then all experts in the
subgroup Gh should decrease the assessment c.1) Preference searching for high-importance sub-
associated with the pair of alternatives ðxi , xj Þ: groups. In such case, fewer changes than the two afore-
b) Preference searching for medium-importance mentioned strategies can be made. The former situation
subgroups. In this situation, the importance of a requires changes for alternatives. Nevertheless, this situ-
subgroup is higher than that of a low-importance ation only requires changes to be made to the pairs of
subgroup. It is reasonable to reduce the number of alternatives whose consensus degrees are smaller than a
P#Gh hr
changes in preference values. In the previous situ- specific proximity threshold cp  ij ¼ r¼1 cpij =#Gh at
ation, all pairs of alternatives that have a low agree- the level of pairs of alternatives.
ment level are required to be modified. In this
situation, the alternatives that have preference values i) For a subgroup Gh 2 Ghigh , the experts whose
with low agreement are selected. An alternative consensus degree is lower than that of this
P
proximity threshold ca  ¼ #Gh cahr =#G is used
i r¼1 i h
subgroup to the global are identified by
here, where cahr
i is the consensus level of alternative EXP 3 ¼ fehr jCLhr < CLðGh ÞjGh 2 Ghigh , ehr 2 Gh g
xi associated with expert er in Gh : Another import-
(25)
ant difference is the number of experts involved in
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY 873

Figure 1. Framework of the proposed model.

ii) For any subgroup Gh 2 Glow , the alternatives 5.1. On the LSGDM evaluation
with a consensus degree lower than threshold
This part selects two typical illustrative examples
c3 are identified by
from Wu and Xu (2018) and Xu, Zhong, et al.
ALT3 ¼ fxi jcai < c3 g (26) (2015) to verify the validity of the index system for
LSGDM problems. These two papers used the same
where c3 can be calculated in analogous to c1 : data set to deal with an emergency large-scale prob-
lem composed of 20 experts.
iii) The pairs of alternatives that need to be modi- Wu and Xu (2018) proposed a consensus model
fied are identified by for LSGDM with changeable clusters based on the
possibility distribution of an HFS. Those 20 experts
POA3 ¼ fði, jÞjcai < c3 Ùxi 2 ALT3 g (27) were divided into 3 clusters with the k-means clus-
tering algorithm. Next, we use our proposed method
iv) The set of preference values selected to make to evaluate their approach. Note that the original
modifications is identified by judgment information is given as fuzzy preference
relations. The results are presented in Table 2.
3 ¼ fði, jÞ 2 POAjcai  
PREhr hr
c a i Ùcphr c p ij g
ij   Xu, Zhong, et al. (2015) introduced an exit-
(28) delegation mechanism to solve LSGDM problems.
They first separated the experts into several clusters,
which are treated as decision units. Then, if the
Generation of advices nearness degree of a cluster is lower than the con-
This situation has the same direction rules as the sensus threshold, this cluster is advised to exit the
preference searching for medium import- decision-making process. However, the “exited”
ance subgroups. cluster can allocate its trust weight to other clusters
A summarisation graph of the proposed model is and delegate them to make decisions. Therefore, the
presented in Figure 1. “exited” cluster still influences the following

Table 2. Evaluation results for the method proposed by Wu


5. Illustrative example and Xu (2018).
In this section, we use two parts to illustrate the Evaluation index

effectiveness and practicality of the two proposed Subgroups Consistency Nearness Evenness
methods. One is the LSGDM evaluation framework, G1 1.000 0.895 0.969
G2 0.875 0.940 0.885
and the other is the adaptive dynamic large- G3 0.833 0.889 0.994
scale CRP. Global group 0.881 0.908 0.951
874 M. TANG ET AL.

Table 3. Evaluation results for the method proposed by Xu, clustering method, the number of clusters and the
Zhong, et al. (2015). weight-determining method for clusters.
Evaluation index
Subgroups Consistency Nearness Evenness
G1 0.708 0.818 0.741
5.2. On the adaptive dynamic consensus
G2 0.833 0.942 0.926 reaching process
G3 0.875 0.906 0.786
G4 0.792 0.902 0.970 Sichuan Province is located in the transition zone
G5 0.833 0.869 0.812
G6 0.854 0.930 0.901
between Qinghai-Tibet Plateau and Sichuan Basin,
Global group 0.815 0.894 0.858 and is frequently affected by topography, geomorph-
ology, rainstorms, earthquakes and other induced fac-
discussions. Based on their model, 20 experts are tors, which make it a geological hazard-prone area. In
classified into 6 clusters. In the following discussion the flood season, Sichuan suffered from multiple
rounds, the first cluster exited the decision process. cycles of heavy rainfall, which caused landslides, col-
Next, we use our proposed method to evaluate their lapses and other geological disasters. To strengthen
approach. The results are shown in Table 3. the prevention of geological hazards in flood season,
From the results of Tables 2 and 3, we can find that Sichuan launched a round of investigation concerning
Wu and Xu (2018)’s method performed better than the hidden dangers of geological hazards with the help
that of Xu, Zhong, et al. (2015). The nearness degree of new technologies such as remote sensing,
of the former study has a slight advantage, while the unmanned aerial vehicles and three-dimensional laser
consistency degree and the evenness degree both have scanning. Suppose that there are four regions
clear advantages. Xu, Zhong, et al. (2015)’s exit-dele- fx1 , x2 , x3 , x4 g in Ya’an City to be selected as candidate
gate mechanism would let a cluster’s experts exit the alternatives. To scientifically assess the disaster risk
decision-making process. Liao, Xu, Zeng, and Xu and take preventive measures in order of priority, 20
(2016) also investigated this issue in the environment experts, represented by E ¼ fe1 , e2 , :::, e20 g, are
of small-scale GDM. They pointed out that the invited to make decisions on these four alternatives.
method that removes experts whose opinions were far Next, we use our proposed model to solve this prob-
from the group opinion had two drawbacks: one was lem. To make comparisons with the previous two
the loss of original judgment information; the other methods, this study uses the same data as Wu and Xu
was attacking the participation motivation of experts. (2018) and Xu, Zhong, et al. (2015). 20 experts pro-
Furthermore, many other factors can also influence vided their evaluation information using fuzzy prefer-
the results of a LSGDM problem, such as the ence relations, which are presented below:

0 1 0 1 0 1 0 1
0:5 0:4 0:4 0:6 0:5 0:3 0:6 0:4 0:5 0:6 0:2 0:3 0:5 0:9 0:7 0:8
B C B C B C B C
B 0:6 0:5 0:3 0:4 C B 0:7 0:5 0:6 0:6 C B 0:4 0:5 0:4 0:3 C B 0:1 0:5 0:8 0:7 C
P1 ¼ B C B
B 0:6 0:7 0:5 0:7 C, P2 ¼ B 0:4
C, P3 ¼ B C, P4 ¼ B C,
@ A @ 0:4 0:5 0:6 C
A
B 0:8
@ 0:6 0:5 0:4 C
A
B 0:3
@ 0:2 0:5 0:1 C
A
0:4 0:6 0:3 0:5 0:6 0:4 0:4 0:5 0:7 0:7 0:6 0:5 0:2 0:3 0:9 0:5
0 1 0 1 0 1 0 1
0:5 0:7 0:4 0:5 0:5 0:9 0:9 0:8 0:5 0:3 0:7 0:8 0:5 0:1 0:6 0:6
B C B C B C B C
B 0:3 0:5 0:1 0:2 C B 0:1 0:5 0:7 0:8 C B 0:7 0:5 0:3 0:6 C B 0:9 0:5 0:6 0:4 C
P5 ¼ B C B
B 0:6 0:9 0:5 0:4 C, P6 ¼ B 0:1
C, P7 ¼ B C, P8 ¼ B C,
@ A @ 0:3 0:5 0:4 C
A
B 0:3
@ 0:7 0:5 0:3 C
A
B 0:4
@ 0:4 0:5 0:3 C
A
0:5 0:8 0:6 0:5 0:2 0:2 0:6 0:5 0:2 0:4 0:7 0:5 0:4 0:6 0:7 0:5
0 1 0 1 0 1 0 1
0:5 0:1 0:8 0:4 0:5 0:4 0:4 0:4 0:5 0:4 0:4 0:2 0:5 0:4 0:4 0:1
B C B C B C B C
B 0:9 0:5 0:6 0:6 C B 0:6 0:5 0:1 0:5 C B 0:6 0:5 0:1 0:2 C B 0:6 0:5 0:5 0:4 C
P9 ¼ B C B
B 0:2 0:4 0:5 0:8 C, P10 ¼ B 0:6
C, P11 ¼ B C, P12 ¼ B C,
@ A @ 0:9 0:5 0:4 C
A
B 0:6
@ 0:9 0:5 0:4 C
A
B 0:6
@ 0:5 0:5 0:7 C
A
0:6 0:4 0:2 0:5 0:6 0:5 0:6 0:5 0:8 0:8 0:6 0:5 0:9 0:6 0:3 0:5
0 1 0 1 0 1 0 1
0:5 0:6 0:4 0:2 0:5 0:6 0:2 0:3 0:5 0:6 0:4 0:1 0:5 0:3 0:6 0:7
B C B C B C B C
B 0:4 0:5 0:3 0:7 C B 0:4 0:5 0:4 0:3 C B 0:4 0:5 0:3 0:4 C B 0:7 0:5 0:8 0:8 C
P13 ¼ B C B
B 0:6 0:7 0:5 0:6 C, P14 ¼ B 0:8
C, P15 ¼ B C, P16 ¼ B C,
@ A @ 0:6 0:5 0:4 C
A
B 0:6
@ 0:7 0:5 0:7 C
A
B 0:4
@ 0:2 0:5 0:9 C
A
0:8 0:3 0:4 0:5 0:7 0:7 0:6 0:5 0:9 0:6 0:3 0:5 0:3 0:2 0:1 0:5
0 1 0 1 0 1 0 1
0:5 0:2 0:6 0:6 0:5 0:4 0:7 0:6 0:5 0:6 0:6 0:7 0:5 0:7 0:6 0:8
B C B C B C B C
B 0:8 0:5 0:8 0:8 C B 0:6 0:5 0:4 0:8 C B 0:4 0:5 0:9 0:9 C B 0:3 0:5 0:6 0:7 C
P17 ¼ B C B C B
B 0:4 0:2 0:5 0:6 C, P18 ¼ B 0:3 0:6 0:5 0:7 C, P19 ¼ B 0:4
C, P20 ¼ B C:
@ A @ A @ 0:1 0:5 0:9 C
A
B 0:4
@ 0:4 0:5 0:9 C
A
0:4 0:2 0:4 0:5 0:4 0:2 0:3 0:5 0:3 0:1 0:1 0:5 0:2 0:3 0:1 0:5
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY 875

fG1 g: The consensus degree at the alternative level


for three subgroups is presented in Figure 3.
a) Preference searching for low-importance
subgroup. For G2 , all experts are required to make
modifications. According to Equations (18) and
(19), alternatives and pairs of alternatives that
should be modified are fx2 , x3 g and fð2, 3Þ, ð2, 4Þ,
ð3, 1Þ, ð3, 2Þg, respectively.

Generation of advices
All experts in subgroup G2 should increase their assess-
Figure 2. The point distribution. ments for positions ð2, 3Þ and ð2, 4Þ, and decrease the
assessments for positions ð3, 1Þ and ð3, 2Þ:
Table 4. The structure of clusters. b) Preference searching medium-importance sub-
Subgroup Gh Number of experts Experts group. For G3 , according to Equations (20)–(22), the
G1 6 e1 , e2 , e8 , e9 , e12 , e17 experts, alternatives and pairs of alternatives are
G2 7 e3 , e5 , e10 , e11 , e13 , e14 , e15 fe4 , e6 , e7 , e16 , e19 , e20 g, fx1 , x2 g and fð1, 4Þ, ð2, 3Þ,
G3 7 e4 , e6 , e7 , e16 , e18 , e19 , e20
ð2, 4Þg, respectively. According to Equation (23),
experts required to modify the selected pairs of alterna-
Table 5. The weight of each sub-group. tives are identified:
Subgroup Gh #Gh Conh Weight
PRE32, 4 ¼ fð1, 4Þ, ð2, 3Þ, ð2, 4Þg,
G1 6 0.938 0.412
G2 7 0.729 0.275 PRE32, 6 ¼ fð1, 4Þ, ð2, 3Þ, ð2, 4Þg,
G3 7 0.792 0.313
PRE32, 7 ¼ fð1, 4Þ, ð2, 3Þ, ð2, 4Þg,
We set d ¼ 0:85, b ¼ 1: Note that this study PRE32, 16 ¼ fð1, 4Þ, ð2, 3Þ, ð2, 4Þg,
assumes that all experts agree to make modifications PRE32, 19 ¼ fð1, 4Þ, ð2, 3Þ, ð2, 4Þg, PRE32, 20 ¼ fð1, 4Þg:
according to the suggestions provided by
the moderator.
First, we use the Euclidean distance to calculate Generation of advices
the distances between these 20 preference relations. The experts in subgroup G3 are required to increase
There are 190 distances altogether. We sort all these the preference value PRE32, 7 ¼ fð2, 3Þg;
distances dij in ascending order and select the trun- The experts in subgroup G3 are required to
cation distance dc to be 0.1958. Figure 2 displays the decrease the following preference values: PRE32, 4 ¼
decision graph with the horizontal axis being fð1, 4Þ, ð2, 3Þ, ð2, 4Þg, PRE32, 6 ¼ fð1, 4Þ, ð2, 3Þ, ð2, 4Þg,
the density of points qi and the vertical axis being PRE23,7 ¼fð1,4Þ,ð2,4Þg, PRE23,16 ¼fð1,4Þ,ð2,3Þ,ð2,4Þg,
the parameter di : Here, according to the Step 5 of PRE23,19 ¼fð1,4Þ,ð2,3Þ,ð2,4Þg, PRE23,20 ¼fð1,4Þg:
Algorithm 1, we set the threshold of qi and di be 5 c) Preference searching high-importance subgroup.
and 0.25, respectively. Based on Figure 2, we can For G1 , according to Equations (24)–(26), experts,
select the appropriate clustering centroids. In this alternatives and pairs of alternatives are
example, we select experts e2 , e14 , e18 to be three fe1 , e8 , e9 , e12 , e17 g, fx1 , x2 g and fð1, 2Þ, ð2, 1Þg,
clustering centroids. Note that expert e3 and e14 are respectively. According to Equation (27), the experts
repeated centroids. We select one of them randomly. who are required to modify the selected pairs of
The structures of three subgroups are shown in alternatives are identified:
Table 4. PRE13, 8 ¼ fð1, 2Þ, ð2, 1Þg, PRE13, 9
After calculating, we obtain the consensus level
0:845 < 0:85, and thus the CRP should be applied. ¼ fð1, 2Þ, ð2, 1Þg, PRE13, 17 ¼ fð1, 2Þ, ð2, 1Þg,
We compute the consistency degree of these three
subgroups and update their weights. The results are
presented in Table 5. Generation of advices
The experts in subgroup G1 are required to increase
Round 1 the following preference values: PRE13, 8 ¼ fð1, 2Þg,
Next, the consensus control process is applied to PRE13, 9 ¼ fð1, 2Þg, PRE13, 17 ¼ fð1, 2Þg:
improve the agreement level. To generate adaptive The experts in subgroup G3 are required to
recommendations, three subgroups are divided into decrease the following preference values: PRE13, 8 ¼
three subsets: Glow ¼ fG2 g, Gmed ¼ fG3 g, Ghigh ¼ fð2, 1Þg, PRE13, 9 ¼ fð2, 1Þg, PRE13, 17 ¼ fð2, 1Þg:
876 M. TANG ET AL.

Round 2 5.3. Comparison with previous two methods


Suppose that the experts receive the modified sug-
To evaluate the effect of our proposed method, we
gestions and provide the following new preferences:
compute three indices: consistency degree, nearness
For experts in subgroup G1 2 Ghigh ,
degree and evenness degree. The results are pre-
sented in Table 6. Furthermore, Table 7 shows some
h812 ¼ 0:3, h821 ¼ 0:7, h912 ¼ 0:2, h921 ¼ 0:8, h17
12
comparisons concerning three methods (our pro-
¼ 0:4, h17
21 ¼ 0:6; posed method; Wu & Xu, 2018; Xu, Zhong,
et al., 2015).
For experts in subgroup G2 2 Glow , The previous two studies used the emergency
problem as an illustrative example. Their consensus
h323 ¼ 0:5, h324 ¼ 0:4, h331 ¼ 0:7, h332 ¼ 0:5; h523 ¼ 0:2, threshold is 0.8. Our illustrative example has a
h524 ¼ 0:3, h531 ¼ 0:5, h532 ¼ 0:8; h10
23 ¼ 0:2, higher consensus threshold of 0.85. Furthermore,
the proposed method used two rounds of iterations
24 ¼ 0:6, h31 ¼ 0:5, h32 ¼ 0:8, h23 ¼ 0:2, h24 ¼ 0:3,
h10 10 10 11 11

and achieved a consensus level of 0.883. From this


31 ¼ 0:5, h32 ¼ 0:8; h23 ¼ 0:4, h24 ¼ 0:7,
h11 11 13 13
point of view, our method performed better, since
31 ¼ 0:5, h32 ¼ 0:8; h23 ¼ 0:5, h24 ¼ 0:4, h31 ¼ 0:7,
h13 13 14 14 14
the final consensus level of the other two methods
32 ¼ 0:5; h23 ¼ 0:4, h24 ¼ 0:5, h31 ¼ 0:5,
h14 15 15 15
was about 0.8. This can occur because of
two reasons:
32 ¼ 0:6;
h15

1. One is the clustering algorithm. Wu and Xu


For experts in subgroup G3 2 Gmed , (2018) used the k-means clustering algorithm
and Xu, Zhong, et al. (2015) applied the vector
h414 ¼ 0:7, h423 ¼ 0:7, h424 ¼ 0:6, h614 ¼ 0:7, h623 ¼ 0:6,
space based clustering algorithm. These two
h624 ¼ 0:7, h714 ¼ 0:7, h723 ¼ 0:4, h724 ¼ 0:5, algorithms belong to the partitioning method.
14 ¼ 0:6, h23 ¼ 0:7, h24 ¼ 0:7, h14 ¼ 0:6, h23 ¼ 0:8,
h16 16 16 19 19 The FSFDP algorithm used in this study is
based on density.
24 ¼ 0:8, h14 ¼ 0:7:
h19 20
2. Another important difference is the weight
determination method. This study took the size
The consensus levels for the three subgroups are: of a subgroup and the consistency degree into
CLðG1 Þ ¼ 0:921, CLðG2 Þ ¼ 0:865, CLðG3 Þ ¼ 0:863: consideration when computing the weight of a
The global consensus level is GCL ¼ 0:883 > 0:85: subgroup. The consistency degree can reflect
Thus, the consensus threshold is achieved. the agreement level in alternative ranking.
Therefore, our proposed method has the best
performance regarding the nearness index. As a
result, the original judgment information can be
reserved as much as possible. A subgroup with
higher consistency degree will be assigned a
larger weight. This can accelerate the CRP.
Furthermore, this integrated weight-determin-
ation method combing two indices is more
comprehensive than single index-based meth-
Figure 3. The consensus degree at the alternative level. ods. For instance, if we only use the scale index,
then the subgroups with the same number of
Table 6. The evaluation results regarding the example.
experts but possessing different inner character-
Evaluation index
istics will have the same weights.
Subgroups Consistency Nearness Evenness
G1 0.938 0.955 0.684
G2 0.729 0.891 0.986
We should note that different studies have their
G3 0.792 0.893 0.987 own emphasis. Wu and Xu (2018) focus on change-
Global group 0.835 0.918 0.863 able clusters. Xu, Zhong, et al. (2015) proposed an

Table 7. The evaluation results of three methods.


Evaluation index
Consensus Number of Final consensus
Methods threshold iterations level Consistency Nearness Evenness
The proposed method 0.85 2 0.883 0.835 0.918 0.863
Wu and Xu (2018) 0.8 3 0.802 0.881 0.908 0.951
Xu, Zhong, et al. (2015) 0.8 3 0.797 0.815 0.894 0.858
JOURNAL OF THE OPERATIONAL RESEARCH SOCIETY 877

exit-delegation mechanism. Our proposed model Disclosure statement


paid attention to the adaptive CRP with hybrid No potential conflict of interest was reported by
strategies for subgroups with different importance. the authors.
It is hard to say which method is the best.

Funding
6. Conclusions
The work was supported by the National Natural Science
Up to now, LSGDM methods developed greatly and Foundation of China (71771156, 71971145).
have been widely researched. However, studies from
some other perspectives, such as how to evaluate the
effects of different LSGDM methods, or how to ORCID
measure the decision results and how to find out
Ming Tang http://orcid.org/0000-0002-4338-699X
each subgroup’s contribution and quality in the Huchang Liao http://orcid.org/0000-0001-8278-3384
decision process, are relatively deficient. Xiaomei Mi http://orcid.org/0000-0001-8651-0033
In this study, we proposed three quality subgroup
indices to measure each subgroup’s influence on the
decision-making process. Next, based on this infor- References
mation, we introduced a method to evaluate the Ding, R. X., Wang, X. Q., Shang, K., & Herrera, F.
results of an LSGDM problem. Furthermore, effects (2019). Social network analysis-based conflict relation-
of different LSGDM methods can also be obtained ship investigation and conflict degree-based consensus
reaching process for large scale decision making using
by comparing their performance under a specific sparse representation. Information Fusion, 50, 251–272.
index. Another important contribution was the pro- doi:10.1016/j.inffus.2019.02.004
posal of the adaptive dynamic CRP. The consistency Farhadinia, B. (2013). A novel method of ranking hesitant
degree of each subgroup was adjusted into the CRP fuzzy values for multiple attribute decision-making
dynamically. Then, we divided the subgroups into problems. International Journal of Intelligent Systems,
28(8), 752–767. doi:10.1002/int.21600
three categories according to their importance.
Herrera-Viedma, E., Cabrerizo, F. J., Kacprzyk, J., &
Three feedback strategies adapting to their weights Pedrycz, W. (2014). A review of soft consensus models
were introduced to improve the consensus level. in a fuzzy environment. Information Fusion, 17, 4–13.
The effectiveness and practicality of our proposed doi:10.1016/j.inffus.2013.04.002
model was illustrated by an example of geological Lahtinen, T. J., H€am€al€ainen, R. P., & Jenytin, C. (2019).
On preference elicitation processes which mitigate the
hazard risk assessment.
accumulation of biases in multi-criteria decision ana-
Based on the comparison work, we found that lysis. European Journal of Operational Research, 282(1),
our consensus model performed better in conver- 201–210. doi:10.1016/j.ejor.2019.09.004
gence speed. The consensus threshold of the pro- Li, C. C., Dong, Y. C., & Herrera, F. (2019). A consensus
posed model was 0.85 while the other two papers model for large-scale linguistic group decision making
reported a threshold of 0.8. Furthermore, our model with a feedback recommendation based on clustered
personalized individual semantics and opposing con-
reached a consensus level of 0.883 after two rounds
sensus groups. IEEE Transactions on Fuzzy Systems,
of iterations. The other two studies reached a con- 27(2), 221–233. doi:10.1109/TFUZZ.2018.2857720
sensus level of about 0.8 after three rounds of itera- Liao, H. C., Xu, Z. S., Zeng, X. J., & Xu, D. L. (2016). An
tions. Several factors can result in this, such as the enhanced consensus reaching process in group decision
weight-determination method for subgroups, clus- making with intuitionistic fuzzy preference relations.
Information Sciences, 329, 274–286. doi:10.1016/j.ins.
tering method or number of clusters.
2015.09.024
In the near future, we will investigate other spe- Liu, B. S., Shen, Y. H., Chen, X. H., Chen, Y., & Wang,
cific factors that may exert influence on LSGDM X. Q. (2014). A partial binary tree DEA-DA cyclic clas-
results, such as the clustering algorithms and the sification model for decision makers in complex multi-
weight-determining methods. In addition, non- attribute large-group interval-valued intuitionistic fuzzy
cooperative behaviour detection and management is decision-making problems. Information Fusion, 18,
119–130. doi:10.1016/j.inffus.2013.06.004
also an important research topic in LSGDM. In the Liu, B. S., Zhou, Q., Ding, R. X., Palomares, I., &
CRP, experts may refuse to accept suggestions or Herrera, F. (2019). Large-scale group decision making
only partially make adjustments. In such case, we model based on social network analysis: Trust relation-
can refer to Lahtinen, H€am€al€ainen, and Jenytin ship-based conflict detection and elimination. European
(2019)’s method to design a preference elicitation Journal of Operational Research, 275(2), 737–754. doi:
10.1016/j.ejor.2018.11.075
process such that the effects of biases counteract
Liu, X., Xu, Y. J., Montes, R., Ding, R. X., & Herrera, F.
each other. In this way, experts are not forced to (2019). Alternative ranking-based clustering and reli-
change their behaviours and make adjustments ability index-based consensus reaching process for hesi-
on judgments. tant fuzzy large scale group decision making. IEEE
878 M. TANG ET AL.

Transactions on Fuzzy Systems, 27(1), 159–171. doi:10. Song, Y. M., & Li, G. X. (2019). A large-scale group decision-
1109/TFUZZ.2018.2876655 making with incomplete multi-granular probabilistic lin-
Liu, Y., Fan, Z. P., & Zhang, X. (2016). A method for large guistic term sets and its application in sustainable supplier
group decision-making based on evaluation information selection. Journal of the Operational Research Society,
provided by participators from multiple groups. Information 70(5), 827–841. doi:10.1080/01605682.2018.1458017
Fusion, 29, 132–141. doi:10.1016/j.inffus.2015.08.002 Tang, M., Zhou, X., Liao, H., Xu, J., Fujita, H., &
Ma, Z. Z., Zhu, J. J., Ponnambalam, K., & Zhang, S. T. Herrera, F. (2019). Ordinal consensus measure with
(2019). A clustering method for large-scale group deci- objective threshold for heterogeneous large-scale group
sion-making with multi-stage hesitant fuzzy linguistic decision making. Knowledge-Based Systems, 180, 62–74.
terms. Information Fusion, 50, 231–250. doi:10.1016/j. doi:10.1016/j.knosys.2019.05.019
inffus.2019.02.001 Torra, V. (2010). Hesitant fuzzy sets. International
Mata, F., Martınez, L., & Herrera-Viedma, E. (2009). An Journal of Intelligent Systems, 25(6), 529–539. doi:10.
adaptive consensus support model for group decision- 1002/int.20418
making problems in a multigranular fuzzy linguistic Wu, Z. B., & Xu, J. P. (2018). A consensus model for
context. IEEE Transactions on Fuzzy Systems, 17(2), large-scale group decision making with hesitant fuzzy
279–290. doi:10.1109/TFUZZ.2005.856561 information and changeable clusters. Information
Nose, S. (1984). A unified formulation of the constant tem- Fusion, 41, 217–231. doi:10.1016/j.inffus.2017.09.011
perature molecular dynamics methods. The Journal of Xia, M. M., & Xu, Z. S. (2013). Managing hesitant infor-
Chemical Physics, 81(1), 511–519. doi:10.1063/1.447334 mation in GDM problems under fuzzy and multiplica-
Palomares, I., Martınez, L., & Herrera, F. (2014). A con- tive preference relations. International Journal of
sensus model to detect and manage noncooperative Uncertainty, Fuzziness and Knowledge-Based Systems,
behaviors in large-scale group decision making. IEEE 21(06), 865–897. doi:10.1142/S0218488513500402
Transactions on Fuzzy Systems, 22(3), 516–530. doi:10. Xu, X. H., Du, Z. J., & Chen, X. H. (2015). Consensus
1109/TFUZZ.2013.2262769 model for multi-criteria large-group emergency deci-
Pang, J. F., & Liang, J. Y. (2012). Evaluation of the results sion making considering non-cooperative behaviors
of multi-attribute group decision-making with linguistic and minority opinions. Decision Support Systems, 79,
information. Omega, 40(3), 294–301. doi:10.1016/j. 150–160. doi:10.1016/j.dss.2015.08.009
omega.2011.07.006 Xu, X. H., Zhong, X. Y., Chen, X. H., & Zhou, Y. J.
Perez, I. J., Cabrerizo, F. J., Alonso, S., Dong, Y. C., Chiclana, (2015). A dynamical consensus method based on exit–-
F., & Herrera-Viedma, E. (2018). On dynamic consensus delegation mechanism for large group emergency deci-
processes in group decision making problems. Information sion making. Knowledge-Based Systems, 86, 237–249.
Sciences, 459, 20–35. doi:10.1016/j.ins.2018.05.017 doi:10.1016/j.knosys.2015.06.006
Perez, I. J., Cabrerizo, F. J., Alonso, S., & Herrera-Viedma, Xu, Y. J., Wen, X. W., & Zhang, W. C. (2018). A two-
E. (2014). A new consensus model for group decision stage consensus method for large-scale multi-attribute
making problems with non-homogeneous experts. IEEE group decision making with an application to earth-
Transactions on Systems, Man, and Cybernetics: Systems, quake shelter selection. Computers & Industrial
44(4), 494–498. doi:10.1109/TSMC.2013.2259155 Engineering, 116, 113–129. doi:10.1016/j.cie.2017.11.025
Rodrıguez, A., & Laio, A. (2014). Machine learning. Xu, Z. S., & Xia, M. M. (2011). Distance and similarity
Clustering by fast search and find of density peaks. measures for hesitant fuzzy sets. Information Sciences,
Science (New York, N.Y.), 344(6191), 1492–1496. doi:10. 181(11), 2128–2138. doi:10.1016/j.ins.2011.01.028
1126/science.1242072 Xu, Z. S., & Zhang, X. L. (2013). Hesitant fuzzy multi-
Rodrıguez, R. M., Labella, A.,  De Tre, G., & Martınez, L. attribute decision making based on TOPSIS with
(2018). A large scale consensus reaching process man- incomplete weight information. Knowledge-Based
aging group hesitation. Knowledge-Based Systems, Systems, 52, 53–64. doi:10.1016/j.knosys.2013.05.011
159(2018), 86–97. doi:10.1016/j.knosys.2018.06.009 Zhu, B., & Xu, Z. S. (2014). Analytic hierarchy process-
Saint, S., & Lawson, J. R. (1994). Rules for reaching con- hesitant group decision making. European Journal of
sensus: A modern approach to decision making. San Operational Research, 239(3), 794–801. doi: 10.1016/j.
Francisco: Jossey-Bass. ejor.2014.06.019.

You might also like