You are on page 1of 15

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

net/publication/375169301

A fair consensus adjustment mechanism for large-scale group decision


making in term of Gini coefficient

Article in Engineering Applications of Artificial Intelligence · November 2023


DOI: 10.1016/j.engappai.2023.106962

CITATION READS

1 30

3 authors, including:

Fanyong Meng
Nanjing University of Information Science & Technology
183 PUBLICATIONS 3,324 CITATIONS

SEE PROFILE

All content following this page was uploaded by Fanyong Meng on 07 November 2023.

The user has requested enhancement of the downloaded file.


Engineering Applications of Artificial Intelligence 126 (2023) 106962

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence


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

A fair consensus adjustment mechanism for large-scale group decision


making in term of Gini coefficient
Fanyong Meng a, b, Dengyu Zhao a, *, Xumin Zhang b
a
The Research Institute for Risk Governance and Emergency Decision-Making, School of Management Science and Engineering, Nanjing University of Information Science
& Technology, Nanjing, 210044, China
b
School of Business, Central South University, Changsha, 410083, China

A R T I C L E I N F O A B S T R A C T

Keywords: With social media and e-democracy development, decision-making environments and problems have become
Large-scale group decision making increasingly complex. Traditional group decision-making, including a small number of decision makers, cannot
Gini coefficient deal with many practical decision-making problems. Large-scale group decision-making has received extensive
Minority opinion
attention as an efficient measure to address this issue. This paper proposes a two-stage optimization consensus
Fairness
Consensus model
mechanism for large-scale group decision-making by the Gini coefficient, which considers both the consensus
among decision makers’ opinions and the fairness of consensus adjustment. First, we introduce an improved
density peak clustering method to cluster decision makers into subgroups that can avoid different clustering
centers being too close. Then, we define the subgroups’ weights by the silhouette coefficient, the subgroup
judgment distance, and their sizes. Respecting the importance of minority opinions, we design a new method to
identify and adjust the weights of the minority opinions, which considers more aspects than previous ones.
Further, we use the Gini coefficient to measure the equity of consensus adjustment of the group and subgroup.
Moreover, we present a two-stage Gini coefficient-based consensus mechanism to obtain the adjusted consensus
individual decision matrices, which can ensure the minimization and fairness of consensus adjustment and
allocation under the set constraints. Based on these results, we give a new large-scale group decision-making
method. Notably, it is the first method that fully considers the opinions of minor subgroups and the fairness
of allocation results. These characteristics are essential in decision-making problems such as resource allocation
and urban public construction. Finally, we show the utility and validity of the new method through a case study
and make the comparative analysis from numerical and principle aspects.

1. Introduction management (Liu et al., 2018b), assessment of lake chiefs (Meng et al.,
2023a), and optimal site selection (Wang et al., 2018). As a branch of
Because of the complexity and uncertainty of the decision-making GDM, LSGDM shows new characteristics (Ding et al., 2020; Xu et al.,
environment, it is challenging to obtain scientific and feasible 2019; Meng et al., 2023b): (i) decision-making problems are more
decision-making results by relying on a single decision maker (DM). complex, (ii) at least 20 DMs are involved in the decision-making pro­
Group decision making (GDM), including multiple DMs, is usually more cess, and (iii) professional background and interest concerns of DMs are
efficient. However, with the continuous development of social media more divergent.
and e-democracy, many real-world decision-making problems require The primary objective of LSGDM is to select the best one from a set of
the participation of a large number of DMs, such as business projects, e- alternatives according to the evaluation information provided by DMs.
democracy, and emergency decision making. In this case, large-scale In LSGDM, more than 20 DMs participate in decision-making and pro­
group decision making (LSGDM), which allows large-scale DMs to vide evaluation information based on several criteria (Liu et al., 2014).
participate in the decision-making process, has been applied to some There are inevitably significant conflicts between the opinions of DMs
decision-making problems, such as water resource management (Gou (Dong and Cooper, 2016). Therefore, it is necessary to apply consensus
et al., 2018), supplier selection (Li et al., 2022a), healthcare analysis to reduce the discrepancies between DMs. As a critical topic in

* Corresponding author.
E-mail address: 202211240009@nuist.edu.cn (D. Zhao).

https://doi.org/10.1016/j.engappai.2023.106962
Received 21 June 2023; Received in revised form 30 July 2023; Accepted 7 August 2023
Available online 28 August 2023
0952-1976/© 2023 Published by Elsevier Ltd.
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

LSGDM, it has attracted many scholars’ attention. For example, Zhang this issue and identified the subgroup with the lowest consensus level
and Meng (2022) developed an interaction feedback consensus mecha­ and few DMs as the minority opinion subgroup. Then, they adjusted the
nism for LSGDM and researched its application in the selection of ideal weight of the minority opinion subgroup by the weight substitution
mobile health apps; Tang et al. (2020) adopted hybrid strategies to method. To avoid overemphasis on minority opinions, Ren et al. (2020)
develop an adaptive consensus reaching process (CRP), in which inter improved Xu et al.’ s method (Xu et al., 2015) and proposed a new
and intra consensus levels of subgroups are measured; Meng et al. method to adjust the weight of minority opinions. To protect minority
(2022) offered an optimal model-based adaptive CRP, where different opinions in LSGDM, Xiao et al. (2022) adjusted the importance of mi­
consensus strategies were employed for various consensus level; by nority opinions through the trust risk coefficient. Nie et al. (2020)
integrating independent and supervised consensus-reaching models, Du defined a non-support degree function to capture the support degree of
et al. (2020) developed a hybrid consensus model to manage minority opinions. Gou et al. (2021) defined the minority opinion sub­
non-cooperative behaviors; Qu et al. (2021) considered the uncertainty group by two characteristics: the lowest consensus level and its cardi­
of unit adjustment cost, and then constructed three mixed integer robust nality below the average level. Currently, the identification of minority
models to achieve consensus; Meng et al. (2023c) viewed the opinions is mainly based on subgroups’ consensus levels and cardinal­
ordinal-cardinal consensus analysis to develop an LSGDM that consid­ ities. It is worth noting that minority opinions have specific reference
ered the uncertain self-confidence of DMs; Zhong et al. (2022) proposed values, but they should not be overemphasized. How to reasonably treat
a non-threshold consensus model, which identified subgroups to be minority opinions is a challenge in LSGDM.
adjusted based on the consensus level, unit consensus cost and adjust­ Numerous psychological studies have demonstrated that equity pref­
ment willingness, and provided adjustment recommendations by an erence is a typical psychological feature of human behaviors, which shows
optimization model; Xiong et al. (2023) discussed three types of that people usually prefer results with higher fairness (Bazarova et al.,
non-cooperative behaviors of subgroups/DMs by the defined cardinal 2012; Cui et al., 2007; Meng et al., 2023d). The ideal decision-making
and ordinal dissimilarity metrics and suggested the weighting penalty results require not only a high degree of consensus but also a high de­
strategy; Tian et al. (2022) offered an improved K-means algorithm and gree of fairness. Liu et al. (2018a) proposed an LSGDM method that
studied the endo-confidence based consensus-reaching by the feedback measures the fairness of alternatives in terms of the distance between the
iteration mechanism; Wan et al. (2022) developed a CRP based on cumulative distributions of subgroups, which is, in fact, the consensus
personalized individual semantic for LSGDM, which guided DMs with analysis for alternatives; Tang and Liao (2021) designed a
unacceptable consensus levels to adjust their preferences via minimum fairness-oriented consensus mechanism considering the fair distribution of
preference adjustment models; Wan et al. (2023) discussed multi-stage changes, where the number of adjusted opinions of each DM is determined
LSGDM with intuitionistic fuzzy information, where the adjusting DMs by its consensus level. However, the method of Tang and Liao (2021) does
were identified according to whether more than one stage is not explicitly define the benchmark of fairness. Opinions of DMs differ in
non-consensus. LSGDM problems, and the amount of opinion adjustment required by
Yu et al. (2022) classified subgroups into three kinds in view of the different DMs to reach consensus requirements also varies. Current
consensus level and trust loss and built a minimum-cost model based on LSGDM methods mainly restrict consideration of the consensus reaching
the differentiated consensus adjustment analysis; Li et al. (2022b) pre­ among DMs while ignoring the fairness in this process. A fair consensus
sented a minimum cost consensus based on robust discrete optimization adjustment mechanism can reduce the non-cooperative behaviors of DMs
to modify the opinions of DMs; Liu et al. (2023a) researched hesitant and promote consensus-reaching efficiency. Further, it also provides an
fuzzy linguistic LSGDM and built an optimization consensus model essential guarantee for the smooth implementation of subsequent
whose goal is to maximize consensus level under the given budget and decision-making results. Therefore, it is necessary to carry out the study of
hesitant bound. Liang et al. (2022) discussed LSGDM in social networks a fair consensus adjustment mechanism.
by taking consensus efficiency as the terminal condition. Concretely, Based on the above analysis, two questions are drawn: (1) How to
when the ratio of consensus improvement level to consensus adjustment treat and handle the minority opinions properly, and (2) When the
amount in a specific iteration is less than the average of all previous consensus is not met, how to ensure the fairness of consensus adjust­
iterations, the consensus iteration is over; Liu et al. (2023b) viewed the ment. To reply these questions, this paper makes the corresponding
bounded confidence of subgroup as a constraint to construct a model to study and develops a new LSGDM method. The significant contributions
maximize the group consensus level and offered a decision process for are summarized as follows:
LSGDM in social networks. Further, Xu et al. (2018) presented a
two-stage CRP to improve consensus level and studied its application to (i) To avoid the issue of the distance between clustering centers
the emergence LSGDM problems. Wang et al., 2021a introduced a being too close, we improve the density peak clustering method
two-stage minimum cost consensus model to improve the group by setting the distance threshold between the DM to be clustered
consensus level. Liang et al. (2023) presented a prospect value consensus and the clustering center;
index for LSGDM with distributed linguistic information and introduced (ii) Besides the consensus level, the identification of minor opinion
an optimization based two-stage LSGDM method. The first stage ob­ subgroups further considers the average cardinality of subgroups
tained the maximum prospect value consensus degree (PVCD) and the and the threshold number of minority opinions;
opinion leader of each subgroup, and the second stage built the opti­ (iii) The weighting update of minor opinion subgroups is determined
mization model to improve the PVCD that took the priority vectors of by the weights of the remaining pro-and anti-subgroups rather
leaders as the consensus adjustment direction. Zhang et al. (2023) also than their numbers, which is more reasonable;
presented a two-stage distributed linguistic LSGDM method aimed at the (iv) The Gini coefficient defines the fairness indices of subgroup and
maximum fairness satisfaction degree and minimum consensus cost. individual consensus adjustments;
Notably, in these two-stage LSGDM methods, only Zhang et al.’s method (v) A two-stage Gini coefficient-based consensus mechanism is
considers the fairness of consensus adjustment. However, it lacks an developed, which can ensure the consensus reaching and stress
objective basis. Further, none of them considers how to deal with mi­ fairness and minor opinion.
nority opinions.
In most decision-making problems, the majority determines the Full consideration of the suggestions of minority groups and the
result, ignoring minority opinions that may contain valuable informa­ fairness of consensus adjustment distribution are essential guarantees
tion. The majority opinions are not necessarily wholly reasonable, and for realizing democratic decision-making results. It provides a reference
minority opinions can provide new perspectives and improve the quality for the government and enterprises to formulate decision-making plans
of decision making (Dreu and West, 2001). Xu et al. (2015) first noted and maximize the objective fairness of decision-making results.

2
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

The rest of the paper is organized as follows: Section 2 provides some Table 1
preliminaries regarding the Gini coefficient and LSGDM. Section 3 in­ International practice classification of the Gini coefficient.
troduces subgroup clustering and consensus measure. Section 4 pro­ Meaning Too More Relatively Big gap Striking
poses a new LSGDM method with a fair consensus adjustment average average reasonable gap
mechanism. Section 5 illustrates the utility of the proposed model Range [0, 0.2] [0.2, 0.29] [0.3, 0.39] [0.4, [0.6, 1]
through a case study. Section 6 provides a comparative analysis to show 0.59]
the superiority of the proposed method. Section 7 draws the conclusion.

2. Preliminaries where qb = ∑Qn b is the proportion of the fair unit, b = 1, 2, …, n. The


a=1
Qa

This section contains two parts. The first part reviews the concept of service unit refers to the allocated resources, such as human, financial,
the Gini coefficient, and the second part lists some basic notations about and material resources. The fair unit refers to the standard used to
LSGDM. measure the equality of resource allocation, such as population and area.

2.1. Gini coefficient 2.2. Knowledge of LSGDM

The Gini coefficient is a relative index defined by Gini (1914) to Let X = {x1 , x2 , …, xm } be the set of alternatives, C = {c1 , c2 , …, cn }
measure the inequality degree of distribution. The most common be the set of criteria, and E = {e1 , e2 , …, el } (l ≥ 20) be the set of DMs.

interpretation of the Gini coefficient is based on the Lorenz curve w = (w1 , w2 , …, wl ) is the weighting vector on E such that lh=1 wh = 1
(Lorenz, 1905), which is an economic curve consisting of the cumulative h h
and 0 ≤ wh ≤ 1 for all h = 1, 2, …, l. Let V = (vij )m×n be the individual
percentage of a country’s total income and the cumulative percentage of
decision matrix offered by the DM eh, where vhij is the assessment of the
the population with that percentage of wealth (see Fig. 1). It is usually
employed to measure the income inequality of residents in a country or alternative xi for the criterion cj defined on [0, u]. Rh = (rijh )m×n is the
region. normalized individual decision matrix such that rijh = vhij /u for all i = 1, 2,
As illustrated in Fig. 1, the horizontal axis of the Lorenz curve is the …, m and all j = 1, 2, …, n. Let G = {g1 , g2 , …, gκ } be the set of subgroups,
cumulative population percentage, and the vertical axis is the cumula­ ∑
and λ = (λ1 , λ2 , …, λκ ) be the weighting vector on G, where κp=1 λp = 1
tive income percentage. The radian of the curve indicates the inequality and 0 ≤ λp ≤ 1 for all p = 1, 2, …, κ. The number of DMs in the subgroup
degree of distribution. The larger the radian is, the more unequal the
gp is denoted as np. Let Rp = (rijp )m×n and Rc = (rijc )m×n be the subgroup
income distribution will be. Conversely, the closer the curve is to the
absolute equality line, the fairer the income distribution is. decision matrix and the group decision matrix, respectively, where rijp =
∑ ∑κ p
Based on the Lorenz curve, the Gini coefficient is defined as (Gini, 1
np
h c
h∈indgp rij and rij = p=1 λp rij for all i = 1, 2, …, m, and all j = 1, 2, …, n,
1914):
and indgp = {l | el Îgp } for all p = 1, 2, …, κ.
SA An LSGDM method usually includes the following five steps (Ding
Gi = (1)
SA + SB et al., 2020): (i) normalizing individual decision matrices, (ii) clustering
subgroups, (iii) consensus analysis, (iv) calculating the ranking values,
where SA and SB are the areas of A and B in Fig. 1, respectively, and GÎ[0, and (v) exploitation, where the second and third steps are the key links.
1]. The closer the Gini coefficient to 0, the more equal the distribution. Zhong et al. (2021) adopted the below formula to measure the
The actual meaning of the Gini coefficient is shown in Table 1. consensus level of subgroup gp,
Considering the calculation complexity of Eq. (1), Mandell (1991)
( ) 1
introduced the below definition to calculate the Gini coefficient. CI gp = 1 − d(Rp , Rc ) (3)
m×n
Definition 1. (Mandell, 1991). Let Sa be the service unit, and Qa be the
∑m ∑n ⃒⃒ p c ⃒⃒
fair unit, a = 1, 2, …, n. The Gini coefficient is defined as: where d(Rp , Rc ) = i=1 j=1 ⃒rij − rij ⃒ is the Manhattan distance between

subgroup decision matrix Rp = (rijp )m×n and group decision matrix Rc =



n ∑
n
|qb Sa − qa Sb | (rijc )m×n , p = 1, 2, …, κ.
(2)
a=1 b>a
Gi = ∑
n Moreover, the group consensus level is defined as the average of
Sa subgroup consensus levels:
a=1

1∑ κ
( )
GCI = CI gp (4)
κ p=1

It is evident that 0 ≤ GCI ≤ 1. The larger the value of GCI, the higher
the consensus degree between subgroups. Let θ be the consensus
threshold. If GCI < θ, we need to improve the consensus level. If
GCI ≥ θ, the selection process is activated.

3. An improved clustering method

This section introduces an improved density peak clustering method


to reduce the dimension of DMs and enhance decision efficiency. In
addition, it discusses the determination of the subgroup weights in terms
of silhouette coefficient, similarity, and cardinality of subgroups.
Clustering aims to divide a set of unlabeled objects into subgroups so
that objects in one subgroup are more similar. To improve the efficiency
of the CRP, various clustering methods are put forward and applied in
Fig. 1. The Lorenz curve. LSGDM, such as K-means clustering method (Lu et al., 2021), fuzzy

3
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

c-means clustering method (Li et al., 2019), similarity measure-based ρ(e1) = 5, ρ(e2) = 5, ρ(e3) = 11, ρ(e4) = 7, ρ(e5) = 11, ρ(e6) = 12, ρ(e7)
clustering method (Gou and Xu, 2021), agglomerative hierarchical = 12, ρ(e8) = 14, ρ(e9) = 12, ρ(e10) = 13, ρ(e11) = 9, ρ(e12) = 8, ρ(e13) =
clustering method (Wang et al., 2018), and density peak clustering 6, ρ(e14) = 3, ρ(e15) = 5, ρ(e16) = 13, ρ(e17) = 6, ρ(e18) = 12, ρ(e19) = 8,
method (Rodriguez and Liao, 2014). In this paper, the density peak ρ(e20) = 12, ρ(e21) = 11, ρ(e22) = 12, ρ(e23) = 11, ρ(e24) = 8, ρ(e25) = 12.
clustering method is adopted, which determines the clustering center Further, based on Eq. (6), the minimum distances of DMs are
based on the below two assumptions (Rodriguez and Liao, 2014): δ(e1) = 1.08, δ(e2) = 0.72, δ(e3) = 0.98, δ(e4) = 0.26, δ(e5) = 0.69,
δ(e6) = 0.73, δ(e7) = 0.87, δ(e8) = 5.57, δ(e9) = 1.24, δ(e10) = 0.24,
(i) The clustering center’s local density is larger than its neighbors. δ(e11) = 1.04, δ(e12) = 0.59, δ(e13) = 0.51, δ(e14) = 0.93, δ(e15) = 0.37,
(ii) The distances between different clustering centers are relatively δ(e16) = 0.22, δ(e17) = 1.27, δ(e18) = 1.39, δ(e19) = 0.7, δ(e20) = 1.33,
large. δ(e21) = 0.01, δ(e22) = 0.77, δ(e23) = 0.3, δ(e24) = 0.15, δ(e25) = 0.1.
Based on the density peak clustering method (Rodriguez and Liao,
The local density of DM eh is defined as (Rodriguez and Liao, 2014): 2014), we can detect 6 clustering centers: e7, e8, e9, e18, e20, and e22,
whose density peaks are 10.44, 77.98, 14.88, 16.68, 15.96 and 9.24,

l
( )
ρ(eh ) = χ dh,u − dc (5) respectively, where dc = 2.
u=1 Thus, the clustering results are G1 = {e7}, G2 = {e6, e8, e10, e16, e23,
{ e25}, G3 = {e9}, G4 = { e11, e12, e13, e14, e15, e17, e18, e19}, G5 = {e13, e20},
1 if x < 0 ∑ ∑n ⃒⃒ h u ⃒⃒ and G6 = {e1, e2, e3, e4, e5, e21, e22, e24}.
where χ (x) = , dh,u = m j=1 ⃒rij − rij ⃒ for all h, u = 1, 2,
0 otherwise i=1
One can find that the distance between the clustering centers e18 and
…, l, and dc is the truncation distance. e20 is d18, 20 = 0.06, which is too close. To break through the limitation,
The minimum distance from eh to eu whose local density is larger we improve the density peak clustering method by introducing the
than ρh is distance threshold β to cluster DMs, summarized as Algorithm 1.
Algorithm 1. An improved density peak clustering method

As the importance of subgroups is different, we now study the


( )
δ(eh ) = min dh,u (6) determination of the subgroup weights in view of the below three as­
u:ρu >ρh
pects: (i) silhouette coefficient, (ii) distance between subgroups, and (iii)
size of subgroups.
for all h = 1, 2, …, l.
The silhouette coefficient is used to measure the quality of clustering
For the DM eh with the highest local density, it defines δ(eh ) =
(Rousseeuw, 1987), and Tang et al. (2020) introduced it into LSGDM to
max(dh,u ).
u define the weights of DMs. Specifically, the silhouette coefficient s (eh) of
However, the distance between the clustering centers obtained by DM eh is defined as:
the density peak clustering method is relatively close, affecting the ⎧
clustering availability. ⎨ 1 − X(eh )/Y(eh ) ifX(eh ) < Y(eh )
s(eh ) = 0 ifX(eh ) = Y(eh ) (7)

Example 1. Suppose that 25 DMs need to be divided into 6 subgroups. Y(eh )/X(eh ) ifX(eh ) < Y(eh )
The initial opinion vector of 25 DMs is o = (1.01, 1.37, 2.50, 2.09, 2.79,
5.22, 5.36, 4.25, 5.73, 4.49, 6.92, 7.51, 8.13, 9.82, 9.26, 4.47, 8.89, where X(eh ) = d(eh , gp ) is the average distance of eh to all DMs in the
5.88, 7.62, 5.82, 3.49, 3.48, 4.79, 2.35, 4.35), namely, each DM offers a subgroup gp, h = 1, 2, …, l, and Y (eh) is the smallest distance of eh to all
real number. DMs in another subgroup gq, namely, Y(eh ) = minκq=1 , q ∕ = pd(eh , gq ).
It is worth noting that − 1 ≤ s(eh ) ≤ 1. The closer the value of s (eh)
According to Eq. (5), the local densities of DMs are
to 1, the better the connection between DM eh and its subgroup.

4
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

Using Eq. (7), the silhouette coefficient of gp is defined as the average (ii) Only one or a few DMs in it.
value of its DMs’ silhouette coefficients, namely,
( ) 1 ∑ To quantify these two aspects, define n = ⌊l /κ⌋ as the average
s gp = s(eh ) (8) number of DMs in each subgroup, where ⌊l /κ⌋ is the largest positive
np eh ∈gp
integer smaller than l/κ. Let α be the proportion threshold of the mi­
Different from the silhouette coefficient, which reflects the closeness nority opinion. In this paper, gp∗ is identified as a minority opinion
between DMs in intra- and inter-subgroups, we further consider the subgroup if and only if CI(gp∗ ) = min{CI(gp )|p = 1, 2, ..., κ} and np∗ ≤
similarity between subgroups by calculating their Manhattan distance. min{n, ⌊l × α⌋}. That’s, the consensus level of gp∗ is the lowest, and its
Specifically, the similarity of subgroup gp for other subgroups is defined cardinality is no larger than n and ⌊l × α⌋, where ⌊l ×α⌋ is the largest
as: positive integer smaller than l× α. Let
{ ⃒ }

( ) 1 ∑ κ
1 ∑ m ∑ n ⃒ ⃒ G = gp ⃒CI(gp ) = min CI(gq ), and np ≤ min{n, ⌊l ×α⌋}, p = 1, 2, ..., κ
∗ ⃒
q=1,2,...,κ
S gp = 1 − ⃒rp − rq ⃒
ij ij (9)
κ − 1 q=1,p∕
=q
m × n i=1 j=1 be the set of minority opinion subgroups.
Step 2: Discussing minority opinions.
for all p = 1, 2, …, κ. Not all minority opinions are essential and need to be considered.
Based on Eqs. (8) and (9), the weight of the subgroup gp is defined as: First, minority opinion subgroups elaborate on the rationality of their
( ) views. Then, all other subgroups thoroughly discuss minority opinions
es(gp ) S gp np and independently make endorsed or negative choices.
λ p = μ1 ∑
κ + μ 2∑
κ ( ) + μ3 ∑
κ (10)
e s(gp )
S gp np Step 3: Modifying the weights of minority opinion subgroups.
p=1 p=1 p=1 For each gp∗ ∈ G∗ :

where np is the cardinality of subgroup gp, p = 1, 2, …, κ, μ1, μ2, and μ3 ∑ ∑


(i) If p∈indGS,p∗ λp − p∈indGO,p ∗ λp > 0, then the importance of minor­
are weighting coefficients, such that μ1, μ2, μ3 ∈[0, 1] and μ1+μ2+μ3 = 1.
ity opinion subgroup gp∗ should be reconsidered, where indGS,p∗
Eq. (10) shows that the weights of subgroups consider three aspects:
and indGO,p are the sub-index sets of endorsed subgroup set GS,p∗

the silhouette coefficient, the similarity between subgroups, and the
and averse subgroup set GO,p for the minority opinion subgroup

number of DMs. The silhouette coefficient reflects the opinion similarity
gp∗ , respectively, such that GS,p∗ ∩ GO,p = ∅ and GS,p∗ ∪ GO,p =
∗ ∗
between DMs in one subgroup to other subgroups, and S (gp) measures
the similarity of different subgroups’ judgments. They belong to G\{p }. Specifically, the weight of gp∗ is updated as:

judgment-based importance indices. The third item in Eq. (10) shows ⎧ ⎧ / ⎫ ⎫


⎨ ⎨ ∑ ⃒⎬ ⎬
the larger the cardinality of one subgroup, the more important it will be. ̂
∑ ∑ ⃒
⃒indG ⃒ , λp∗
S,p∗
λ p∗ = max min λ − λp , λp
It is an important index based on the number of DMs in subgroups. ⎩ ⎩p∈indGS,p∗ p O,p ∗ p∈indGS,p∗
p∈indG
⎭ ⎭

(11)
4. A new LSGDM method
⃒ ⃒
where ⃒indGS,p∗ ⃒ is the cardinality
⃒ of S,p∗
⃒ indG .
Ideal decision-making results should have not only a high degree of ∑
The term p∈indGS,p∗ λp /⃒indGS,p∗ ⃒ in Eq. (11) is offered to avoid over­
consensus but also a high degree of fairness. To achieve this goal, emphasizing the importance of the minority opinion subgroup gp∗ ,
choosing a rational index to measure the fairness of the consensus which is no greater than the average weight of its endorsed subgroups.
adjustment is necessary. With these considerations, this section proposes
a two-stage Gini coefficient based LSGDM method regarding minority (ii) If
∑ ∑
p∈indGS,p∗ λp − p∈indGO,p ∗ λp < 0, delete the minority opinion
opinion and fairness. subgroup gp∗ due to its minor weight and hindering the consensus
improvement.
∑ ∑
4.1. Managing minority opinions (iii) If p∈indGS,p∗ λp − p∈indGO,p ∗ λp = 0, its weight is unchanged

After clustering, there may be subgroups with small cardinality, and According to the above process, the normalized weights of subgroups
the opinions of DMs in these subgroups are far from those of others, are
known as the minority opinion (Xu et al., 2015). Since such subgroups’
consensus levels and weights are relatively small, minority opinions are ̂
λ p∗
⎧ λ′p∗ = ∑ ∑ p∗ ∈ indGI∗
usually overlooked (Ren et al., 2020). However, the majority opinions ⎪

⎪ λq + ̂
λ p∗

are not those of all, and their views are not always reasonable. The ⎪

⎪ q∈indG\G∗ p∗∈indG∗

principle of “minority following majority” may affect the quality of the ⎨
λ′p = 0 p ∈ indGII∗ (12)
final decision-making results. Although minority opinion is only a small ⎪


part of all judgment information, it may determine the direction of ⎪


⎪ ′ λq { }
decision-making problems. Minority opinions may contain valuable in­

⎩ λp = ∑ ∑ p ∈ indG\ GI∗ , GII∗
λq + ̂
λ p∗
formation and provide a different perspective on the decision-making q∈indG\{GI∗ ,G∗II } p∗∈indGI∗
process (Bazarova et al., 2012). A minority may inspire the majority
to re-examine the decision-making problem and thus generate different where indGI∗ and indGII∗ are the sub-index sets of
ideas (Bazarova et al., 2012). Therefore, we should fully consider the ⃒∑
⃒ ∑
G = {gp ⃒ p∈indGS,p∗ λp −
I∗
p∈indGO,p λp ≥ 0, gp ∈ G }

and G II∗
=
opinions of all DMs and adequately treat and handle the minority

⃒∑ ∑
opinion to get more reasonable and comprehensive decision results. ⃒
p∈indGO,p ∗ λp < 0, gp ∈ G }, respectively.

{gp ⃒ p∈indGS,p∗ λp −
A new method is proposed to treat minority opinions properly, which
comprises three phases: identification, discussion, and modification. Example 2. Suppose that there are six subgroups gp , p = 1, 2, ..., 6, in
Step 1: Identifying minority opinions. an LSGDM problem. The weighting vector on the subgroup set is λ =
Subgroups identified as a minority should satisfy the following two (0.2, 0.1, 0.05, 0.15, 0.24, 0.26). g2 is identified as a minority opinion
conditions: subgroup. After a full discussion, subgroups g1, g3, g4, and g5 support the
opinion of g2, while subgroup g6 opposes it.
(i) The consensus level of the subgroup is the lowest;

5
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

Adopting the new method, the updated weight of the minority level, part subgroups should adjust their opinions. Suppose that the
opinion subgroup g2 is modified opinion vector of subgroups is o = (0.20, 0.4, 0.5, 0.4, 0.6, 0.4),
where the corresponding group consensus is GCI = 0.91, which is larger
̂
λ 2 = max{min{0.64 − 0.26, 0.16}, 0.1} = 0.16 than 0.85. To measure the fairness of this consensus adjustment result by
Further, by Eq. (12), the weighting vector on the subgroup set is Eq. (13), the below calculation process is needed:

(i) Calculating the service units of subgroups, it derives S1 = 0.10,



λ = (0.19, 0.15, 0.04, 0.14, 0.23, 0.25)
In this example, when Xu et al.’s method (Xu et al., 2015) is adopted, S2 = 0.20, S3 = 0.10, S4 = 0.10, S5 = 0.10, and S6 = 0.00;
(ii) Calculating the fair units of subgroups, it gets Q1 = 0.33, Q2 =
the adjusted weight vector on the subgroup set is
0.23, Q3 = 0.16, Q4 = 0.13, Q5 = 0.27, and Q6 = 0.03
λ″ = (0.18, 0.18, 0.04, 0.14, 0.22, 0.24)
According to Eq. (13), we have Gi = 0.25, which shows a high
where the weight of the minority opinion subgroup g2 is 0.18. Specially, fairness degree of subgroup consensus adjustments.
the first and second subgroups have the same weight. This example According to Eq. (2), the fairness index of consensus adjustment of
shows that both methods can increase the importance of minority DMs in subgroup gp is defined as:
opinion subgroups. Compared with Xu et al.’s method (Xu et al., 2015), ∑ ∑ p p p p
the increasing degree of the new method is smaller. ( ) h∈indgp ,h>k |qh Sk − qk Sh |
(14)
k∈indgp
Gi gp = ∑ p
h∈indgp Sh

4.2. Fairness measure based on Gini coefficient ∑ ∑n ⃒⃒ p ⃒


for all p = 1, 2, …, κ, where Sph = m i=1
p ⃒
j=1 ⃒rh,ij − rh,ij ⃒ is the service unit,

The Gini coefficient is an important indicator to measure the fairness ∑ ∑n ⃒⃒ p ⃒ p


Qph = m i=1
c⃒
j=1 ⃒rh,ij − rij ⃒ is the fair unit, and qh =
p ∑ Qh p is the pro­
Q
of income distribution, which is widely used in resource distribution (Ma k∈indgp k
p
et al., 2020; Sueyoshi et al., 2021). Based on the Gini coefficient, this portion of the fair unit for all hÎindgp, Rh = (rph,ij )m×n is the adjusted in­
subsection studies the fairness of consensus adjustments inter- and dividual decision matrix for Rph = (rh,ij
p
)m×n , and Rc = (rijc )m×n is the group
intra-subgroups in the CRP. The consensus adjustment of each
decision matrix.
DM/subgroup is regarded as the service unit of the Gini coefficient. The
distance between individual/subgroup’s decision matrix and group de­ Example 4. Suppose there are three DMs eh , h = 1, 2, 3, in subgroup
cision matrix is deemed as the fair unit of the Gini coefficient. The larger g1 in Example 3, and their weights are equal. The initial opinion
the distance between individual/subgroup’s decision matrix and group vector of DMs in g1 is o = (0.1, 0.2, 0.6). Their adjusted opinion
decision matrix, the larger the required consensus adjustment should be. vector is o = (0.6, 0.6, 0.7). Similar to Example 3, one can get
According to the above analysis, the fairness index of subgroup Gi(g1 ) = 0.405 by Eq. (14). The division standard of fairness in
consensus adjustment is defined as: Table 1 shows a low degree of fairness of consensus adjustment of
∑ ∑ DMs in g1 .
|qr Sp − qp Sr |
(13)
p∈indG\GII∗ II∗
Gi = ∑r∈indG\G ,r>pq
q∈indG\GII∗ S 4.3. Two-stage Gini coefficient based consensus mechanism
∑ ∑n ⃒⃒ p p ⃒⃒
where Sp = m i=1 j=1 ⃒rij − rij ⃒ is the service unit, The consensus adjustment will be activated when the group
p
∑m ∑n ⃒⃒ p c ⃒⃒ p Q p consensus level is insufficient. To retain original information as
Q = i=1 j=1 ⃒rij − rij ⃒ is the fair unit, and q = ∑ is the fair
Qq q∈indG\GII∗ much as possible and to achieve the consensus adjustment as fair as
p
unit proportion for all q ∈ indG\GII∗ , R = (rpij )m×n is the adjusted sub­ possible, this subsection proposes a two-stage Gini coefficient based
consensus mechanism: minimum adjustment consensus and
group decision matrix for Rp = (rijp )m×n , and Rc = (rijc )m×n is the group
maximum fairness consensus. Especially, stage 1 determines the
decision matrix.
∑ optimally adjusted decision matrix of each subgroup by minimum
It is evident that 0 ≤ Gi ≤ 1 because of r∈indG\GII∗ ,r>p |qr Sp − qp Sr | ≤
adjustment consensus model, and stage 2 ascertains the optimally
Sp , q ∈ indG\GII∗ . The smaller the value of Gi, the higher the fairness of adjusted decision matrix of each DM in each subgroup, respectively.
r q
the consensus adjustment. Concretely, when Gi = 0, it acquires QSr = QSq In this process, the fairness of consensus adjustment is stressed by
for all p, r ∈ indG\GII∗ such that p ∕= r. It shows that each subgroup’s Eqs. (13) and (14).
consensus adjustment is proportional to the Manhattan distance be­
tween its decision matrix and group decision matrix. Eq. (3) indicates 4.3.1. Stage 1: Gini coefficient-based consensus adjustment models for
when Gi = 0, each subgroup’s consensus adjustment is proportional to subgroups
its consensus level. Thus, Gi = 0 shows the full fairness, and Gi = 1 The value of the minority opinion lies in its difference from the
means that the smallest fair unit bears all service units. Concretely, the majority opinions, and too much adjustment will cause it to lose this
subgroup with the smallest distance from the group decision matrix is characteristic. To maintain this merit of the minority opinion, this
allocated the whole consensus adjustment. This case corresponds to the subsection limits the adjustment scope of minority opinions through
most unfairness. supporting subgroups. Specifically, the larger the sum of the supporting
subgroups’ weights is, the less the consensus adjustment of the minority
Example 3. (i.g., Example 2) Suppose that there are six subgroups gp ,
opinion subgroup will be. In terms of this principle, the consensus
p = 1, 2, ..., 6, in an LSGDM problem. The weighting vector on the sub­
adjustment of each subgroup gp∗ ∈ G∗ is restricted as:
group set is λ = (0.15, 0.05, 0.15, 0.15, 0.24, 0.26). Let θ = 0.85, and the
⎛ ⎞
initial opinion vector of subgroups be o = (0.1, 0.2, 0.6, 0.3, 0.7, 0.4), ⃒ p∗ ⃒ ∑
then the group opinion is oc = 0.423. By Eq. (3), the consensus levels of ⃒rij − rp∗ ⃒
ij ≤
⎝1− λp ⎠ × rijp∗ (15)
subgroups are p∈indGS,p∗

CI(g1 ) = 0.67, CI(g2 ) = 0.77, CI(g3 ) = 0.83, CI(g4 ) = 0.87, CI(g5 ) for all i = 1, 2, …, m, and all j = 1, 2, …, n, where rijp∗ and rp∗
ij are the
= 0.73, CI(g6 ) = 0.97
original and adjusted judgments of the minority opinion subgroup gp∗ .
Eq. (4) shows that GCI = 0.81 < 0.85. To improve the consensus The smaller the value of Eq. (13) is, the higher the fairness of the

6
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

consensus adjustment among subgroups will be. However, complete ∑ ∑m ∑n p


fairness is difficult or impractical. As shown in Table 1, the fairness min i=1
ε
j=1 ij
degree belonging to [0.2, 0.39] is good. Therefore, we restrict the
p∈indG∗ \GII∗

fairness of consensus adjustment as follows: ∑


∑ ∑ S p = Λ∗
|qr Sp − qp Sr | p∈indG\GII∗
(16) ⎧
p∈indG\GII∗ II∗
0.2 ≤ Gi = ∑r∈indG\G ,r>pq ≤ 0.39 ∑ ∑
q∈indG\GII∗ S



⎪ |qr Sp − qp Sr |



⎪ p∈indG\GII∗ r∈indG\GII∗ ,r>p

where indG\GII∗ is the subscript set of subgroups G by deleting the mi­ ⎪


∑ ≤ 0.39
⎪ Sq
nority opinion subgroups GII∗ , as shown in Eq. (12). ⎪


⎪ q∈indG\GII∗

The purpose of the adjustment is to achieve the consensus require­ ⎪

⎪ ∑ ∑

ment. Therefore, the consensus level of adjusted subgroup opinions ⎪



|qr Sp − qp Sr |

should meet the following condition: ⎪

⎪ 0.2 ≤
p∈indG\GII∗ r∈indG\GII∗ ,r>p




⎪ Sq
GCI ≥ θ (17) ⎪


⎪ q∈indG\GII∗


⃒ ⃒ ⎪
⎪ ( )
where GCI = G\G 1

CI(gp ) with ⃒G\GII∗ ⃒ being the cardinality


⎪ 1 ∑ 1 ∑m ∑n ⃒⃒ p c⃒

| II∗ | p∈indG\GII∗ ⎪
⎪⃒
⎪ ⃒ ⃒ 1− ⃒r − rij ⃒ ≥ θ
j=1 ij
∑m ∑n ⃒⃒ p c ⃒⃒ ⎪
⎪ G\GII∗ ⃒p∈indG\GII∗ m × n i=1
1 h h ⎨
of G\GII∗ , and CI(gp ) = 1 − m×n j=1 ⃒rij − rij ⃒ such that R = (rij )m×n
i=1 s.t. ⎛ ⎞ (19)
c ⎪
is the adjusted subgroup decision matrix, and R = (rcij )m×n is the ⎪

⎪ ⃒ p∗ ⃒
⃒rij − rij ⃒ ≤ ⎝1 −
p∗




⎪ λp ⎠ × rijp∗ + εp∗ ∗ ∗
ij , p ∈ indG \G
II∗

adjusted group decision matrix with rcij = p∈indG\GII∗ λ′p rpij for all i = 1, 2, ⎪



p∈indG S,p∗


…, m, and all j = 1, 2, …, n. ⎪


⎪ εp∗
ij ≥ 0, gp∗ ∈ G \G
∗ II∗
Combining Eqs. (15)–(17), the following minimum adjustment ⎪


⎪ ∑
⎪ c
consensus model is constructed:
κ

⎪ rij = λ′ rp

⎪ p=1 p ij
∑ ⎪


⎪ ∑m ∑n ⃒ p ⃒
min Sp ⎪
⎪ Sp =
⎪ ⃒r − rpij ⃒
p∈indG\GII∗

⎪ i=1 j=1 ij



⎪ p
∑ ∑ ⎪
⎪ rij ∈ [0, 1]


|qr Sp − qp Sr | ⎪


⎧ ⎪
⎪ p ∈ indG\GII∗
p∈indG\GII∗ r∈indG\GII∗ ,r>p ⎪

⎪ ∑ ≤ 0.39 ⎪




⎪ Sq i = 1, 2, ..., m



⎪ q∈indG\GII∗

⎪ ∑ ∑

⎪ j = 1, 2, .., n

⎪ |qr Sp − qp Sr |



⎪ where the first constraint derives model (18) by deleting the fourth
p∈indG\G r∈indG\GII∗ ,r>p
II∗

⎪ 0.2 ≤ ∑

⎪ Sq constraint, εp∗ p∗
ij is a slack variable for the minority subgroup opinion rij ,





⎪ q∈indG\GII∗


⎪ ( ) and all other notations as listed in model (18).




⎪ 1 ∑ 1 ∑m ∑n ⃒⃒ p c⃒

⎪ ⃒ ⃒ 1− ⃒r − rij ⃒ ≥ θ


⎪ ⃒G\GII∗ ⃒ m × n i=1 j=1 ij 4.3.2. Stage 2: Gini coefficient based consensus adjustment models for DMs
⎪ p∈indG\GII∗


⎛ ⎞ From stage 1, we can obtain the consensus adjustment of each sub­
s.t. ⃒ ⃒ ∑ (18) group. The next question is how to allocate each subgroup consensus
⎪ ⃒ p∗

adjustment to its DMs. Let S∗p be the consensus adjustment of subgroup
p∗ ⃒


⎪ rij − rij ≤ ⎝1 − λp ⎠ × rijp∗ , p∗ ∈ indG∗ \GII∗


p ∈ indG\GII∗ derived from model (18) or (19).
S,p∗

⎪ p∈indG

⎪ ∑κ p
On the other hand, the subgroup decision matrices R = (rpij )m×n ,



⎪ rcij = λ′ rp

⎪ p=1 p ij



⎪ ∑m ∑n ⃒ p ⃒ p ∈ indG\GII∗ , should meet the consensus requirement, namely, the

⎪ p ⃒r − rpij ⃒ ∑ p
⎪S =
⎪ rk,ij

⎪ i=1 j=1 ij
fourth constraint in model (19) should hold, where rpij = k∈indg p
for all


⎪ rp ∈ [0, 1] |gp |

⎪ ⃒ ⃒
⃒ ⃒
⎪ i = 1, 2, …, m, all j = 1, 2, …, n, and ⃒gp ⃒ is the cardinality of gp.
ij




⎪ p ∈ indG\GII∗



⎪ Table 1 shows that the smaller the Gini coefficient, the higher the


⎩ i = 1, 2, ..., m fairness. Therefore, we further build the below model to determine the
consensus adjustment of each DM:
j = 1, 2, .., n ∑ ∑ p p p p
( ) k∈indgp h∈indgp ,h>k |qh Sk − qk Sh |
minp∈indG\G Gi gp = ∑
where the first four constraints are derived from Eqs. (15)–(17), the fifth
II∗ p
h∈indgp Sh
constraint offers the calculation of the collective judgment with λ′p being
shown in Eq. (12), indG\GII∗ is the sub-index set of G\GII∗ , the sixth
constraint is the total adjustment of subgroup decision matrix Rp =
(rijp )m×n , and the seventh constraint limits the range scope of adjusted
subgroup judgment.
By solving model (18), the consensus adjustment of each sub­
group decision matrix can be obtained. However, model (18) may be
infeasible. In this case, we remove the fourth constraint condition in
model (18) and resolve it. Let Λ* be the optimal value of model (18)
by deleting the fourth constraint condition. Then, we further
construct the below model to adjust subgroup decision matrices:

7
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

⎧ ∗p ∑m ∑n ⃒ p ⃒
⎪ S = ⃒r − rpij ⃒ ∑ ∑

⎪ i=1 j=1 ij min Skp



⎪ ( ) p∈indG\GII∗ k∈indgp

⎪ 1 ∑ 1 ∑m ∑n ⃒⃒ p ⃒

⎪ ⃒ ⃒ c ⃒
⎪ 1− ⃒r − rij ⃒ ≥ θ


⎪ ⃒G\GII∗ ⃒ m × n i=1 j=1 ij ⎧ ∑ ( )

⎪ p∈indG\GII∗
⎪ Gi gp = Gi∗

⎪ ⎪


⎪ ∑κ ⎪
⎪ p∈indG\G II∗

⎪ rcij = λ′ rp ⎪


⎪ p=1 p ij ⎪ ∑m ∑n ⃒
⎪ ⃒

⎪ ⎪
⎪ ⃒ p p ⃒ p

⎪ / ⎪
⎪ ⃒r − rk,ij ⃒ = Sk
j=1 k,ij

⎪ ∑ p ⃒ ⃒ ⎪
⎪ i=1

⎪ ⎪
⎪ ∑ ∑
⎪ p
⎪ rij =
⎪ rk,ij ⃒gp ⃒ ⎪

⎨ ⎪

⎪ |qp Sp − qpk Shp |
k∈indgp ⎪ ( ) k∈indgp h∈indgp ,h>k h k
s.t. (20) ⎪

⎪ ∑ p

⎪ ∑m ∑n ⃒⃒ p p ⃒
⃒ ⎪
⎪ Gi gp =

⎪ Skp = ⃒r − r ⃒ ⎪
⎪ Sh

⎪ i=1 j=1 k,ij k,ij ⎪


⎪ ⎪
⎪ h∈indgp

⎪ ⎪

⎪ r ∈ [0, 1]
⎪ p ⎪
⎪ ∑m ∑n ⃒ p ⃒

⎪ k,ij ⎪ S∗p =
⎪ ⃒r − rpij ⃒

⎪ ⎪
⎪ i=1 j=1 ij

⎪ ⎪


⎪ k ∈ indgp ⎪
⎪ ( )
⎪ ⎪





⎪ 1 ∑ 1 ∑m ∑n ⃒⃒ p c⃒


⎪ p ∈ indG\GII∗ ⎪
⎪ ⃒ ⃒ 1 − ⃒r − r ⃒ ≥θ

⎪ ⎨ ⃒G\GII∗ ⃒ m × n i=1 j=1 ij ij



⎪ s.t. p∈indG\GII∗
(22)

⎪ i = 1, 2, ..., m ⎪
⎪ ∑κ
⎪ ⎪

⎩ ⎪

⎪ rcij = λ′ rp
p=1 p ij
j = 1, 2, .., n ⎪


⎪ /

⎪ ∑ p

⎪ ⃒ ⃒
where the first constraint defines the consensus adjustment of subgroup ⎪ rpij =

⎪ r k,ij
⃒gp ⃒


p ∈ indG\GII∗ , the fourth constraint shows the relationship between ⎪


k∈indgp

subgroup and individual adjustments, the fifth constraint is the total ⎪

⎪ p
∑m ∑n ⃒⃒ p p ⃒


consensus adjustment of individual decision matrix Rpk , the sixth ⎪ Sh =

⎪ i=1
⃒r − rk,ij ⃒
j=1 k,ij



constraint defines the range scope of adjusted individual judgments, and ⎪ p

⎪ r ∈ [0, 1]

⎪ k,ij
all other notations as shown in model (19). ⎪



The objective function of model (20) makes the Gini coefficient as ⎪ k ∈ indgp



small as possible. Because of the independence of DMs in different ⎪
⎪ p ∈ indG\GII∗



subgroups, we can further transform model (20) into the following ⎪

model: i = 1, 2, ..., m, j = 1, 2, .., n
∑ ( )
min Gi gp where the first constraint comes from the objective function of model
p∈indG\GII∗ (21), the second constraint defines the consensus adjustment of DMs,
∑ ∑ and all other notations are the same as those in model (21).



|qph Skp − qpk Shp | By addressing model (22), we can derive the adjustments of DMs,
⎪ ( )



⎪ Gi gp =
k∈indgp h∈indgp ,h>k
∑ p denoted by S∗pk for all p ∈ indG\G
II∗
and all k ∈ indgp . Under the ascer­




Sh tained individual total consensus adjustment, we further build the below

⎪ model to derive individually adjusted decision matrices:
h∈indgp



⎪ ∑m ∑n ⃒ p ⃒

⎪ ⃒r − r p ⃒ ∑ ∑ ∑m ∑n ⃒⃒ p ⃒

⎪ S∗p = j=1 ij ij ⃒

⎪ i=1 min ⃒r − rcij ⃒

⎪ ( ) i=1 j=1 k,ij

⎪ p∈indG\GII∗ k∈indgp


⎪ 1 ∑ 1 ∑m ∑n ⃒⃒ p c⃒


⎪ ⃒ ⃒ 1 − ⃒r − r ⃒ ≥θ ⎧

⎪ ⃒G\GII∗ ⃒p∈indG\GII∗ m × n i=1 j=1 ij ij
( )

⎪ ⎪





⎪ 1 ∑ 1 ∑m ∑n ⃒⃒ p c⃒


⎪ ∑κ ⎪
⎪ ⃒ ⃒ 1 − ⃒r − r ⃒ ≥θ
j=1 ij
λ′ rp ⎪ ⃒G\GII∗ ⃒p∈indG\GII∗
⎪ ij
⎪ c
⎨ rij = m × n i=1
p=1 p ij ⎪

s.t. / (21) ⎪

⎪ ∑κ

⎪ ⎪


⎪ ∑ p ⃒ ⃒ ⎪ c
⎪ rij =
⎪ λ′ rp
p=1 p ij

⎪ r p
= r ⃒gp ⃒ ⎪
⎪ ij
⎪ k,ij ⎪
⎪ /

⎪ k∈indgp ⎪


⎪ ⎪
⎪ ∑ p ⃒ ⃒

⎪ ∑m ∑n ⃒⃒ p ⃒ ⎪
⎪ r p
= r ⃒gp ⃒

⎪ p ⃒

⎪ ij k,ij
⎪ Shp =
⎪ ⃒r − rk,ij ⃒ ⎪
⎨ k∈indgp
⎪ j=1 k,ij
⎪ (23)
i=1

⎪ s.t.


⎪ p ⎪ ∗p ∑m ∑n ⃒⃒ p
⎪ p ⃒

⎪ rk,ij ∈ [0, 1]




⎪ Sk = i=1 j=1
⃒rk,ij − rk,ij ⃒

⎪ ⎪


⎪ k ∈ indgp ⎪ p

⎪ ⎪

⎪ rk,ij ∈ [0, 1]

⎪ ⎪

⎪ ⎪

⎪ p ∈ indG\GII∗ ⎪


⎪ ⎪
⎪ k ∈ indgp

⎩ ⎪



i = 1, 2, ..., m, j = 1, 2, .., n ⎪


⎪ p ∈ indG\GII∗




where the first constraint acquires from the objective function of model ⎩ i = 1, 2, ..., m, j = 1, 2, .., n
(20), and all other notations are the same as those in model (20).
By addressing model (21), we can derive the minimum value of the where the fourth constraint defines the consensus adjustment of DM ek for all
Gini coefficient, which shows the fairness level of consensus adjustments k ∈ indgp and all p ∈ indG\GII∗ , and all other notations as shown in model
among DMs as high as possible. Without loss of generality, let Gi* be the (22). The objective function of model (23) defines the total deviation be­
optimal objective function value of model (21). We further build the tween adjusted individual judgment and corresponding group judgment.
following model to determine the individual total consensus adjustment:

8
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

4.4. An algorithm DMs, including managers, experts, and the elderly, denoted as E = {e1,
e2, …, e25}. Five criteria are selected to evaluate alternatives: environ­
In view of the above discussion, this subsection offers a new LSGDM ment (c1), technology (c2), service (c3), security (c4), and operation (c5),
method, summarized in Algorithm 2. whose weights are 0.20, 0.15, 0.25, 0.30, and 0.10, respectively. After
preliminary screening, there are four alternatives X = {x1, x2, x3, x4},
Algorithm 2. A new LSGDM method based on the Gini coefficient.
which are

5. A case study x1: Cooperate with surrounding medical institutions.


x2: Outsource medical services to professional medical institutions.
With the global escalation of the aging process, the pension and x3: Set up corresponding medical facilities and departments in the
medical care needs of the elderly are increasing. Currently, nursing nursing home.
homes are far from meeting the medical service needs of the elderly, x4: Build a comprehensive service organization by integrating med­
which makes the social security and health care system face unprece­ ical and elderly care.
dented pressure in China. In this case, the combination of medical and Based on these criteria, each DM needs to evaluate the four alter­
elderly care, which integrates modern medical service technology with natives using 1–10 scale. The initial individual decision matrices Vh (h =
the traditional old-age security model to provide professional medical 1, 2, …, 25) are shown in Table 2.
care services for the elderly, has been widely promoted. It can effectively Step 1: Cluster all DMs into five subgroups based on Algorithm 1 and
integrate resources to meet the diverse needs of the elderly and to Eq. (10). The results are shown in Table 3.
improve the service quality of nursing homes. Step 2According to Table 3, the subgroup decision matrices are ob­
A nursing home in Changsha transforms into a medical and nursing tained as follows:
institution to adapt to market variations. This nursing home invites 25

Table 2
Initial individual decision matrices.
c1 c2 c3 c4 c5 c1 c2 c3 c4 c5

e1 x1 2 10 8 2 10 e2 x1 4 1 7 7 8
x2 7 6 4 3 10 x2 3 8 4 2 1
x3 3 3 4 8 9 x3 9 8 4 9 2
x4 2 9 2 6 6 x4 8 4 7 1 8
e3 x1 1 1 7 9 9 e4 x1 3 10 10 1 6
x2 5 10 6 4 1 x2 8 6 1 3 9
x3 10 7 5 10 3 x3 3 1 6 6 4
x4 3 7 5 1 9 x4 1 6 4 9 3
e25 x1 7 8 2 6 6
x2 1 9 9 10 4
x3 10 1 1 8 1
x4 6 1 6 1 10

Let the truncation distance, the distance threshold, the weight coefficients, the consensus threshold, and the threshold of the size of the minority opinion subgroup be
defined as: dc = 3.50, β = 5.00, μ1 = 0.30, μ2 = 0.30, μ3 = 0.40, θ = 0.85, and α = 0.15, respectively. To obtain the most suitable alternative, the following steps are
applied.

9
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

Table 3 are.
Clustering result and the weight of subgroup. S1 = 2.45, S2 = 0.18, S3 = 2.96, S4 = 1.33, S5 = 1.00.
Subgroup DM Silhouette Distance Sizes of Weights of Step 5: Using models (21) and (22), the consensus adjustments of
coefficient between subgroup subgroup DMs are
subgroup ⎧

⎪ S21 = 2.16, S31 = 2.30, S81 = 2.33, S11
1 1
= 1.96, S12 1
= 2.72, S13 = 3.20;
g1 e2, e3, 0.40 0.32 6 0.21 ⎪


⎪ 2 2 2 2
e8, e11, ⎪ S
⎪ 7 = 0.18, S17 = 0.15, S 21 = 0.19, S 25 = 0.17;
e12, e13,

⎨ 3
S1 = 3.11, S43 = 3.35, S63 = 2.64, S93 = 2.57, S316 = 2.75, S20
3
g2 e7, e17, 0.49 0.37 4 0.18 3
⎪ = 3.40, S23 = 2.93;
e21, e25 ⎪


g3 e1, e4, 0.42 0.34 7 0.23

⎪ S54 = 1.45, S10
4
= 1.29, S144
= 1.26;


e6, e9, ⎪
⎩ S5 = 1.05, S5 = 1.15, S5 = 0.85, S5 = 0.98, S5 = 0.96.
e16, e20,
15 18 19 22 24

e23
Step 6: By model (23), we can obtain the individually adjusted de­
g4 e5, e10, 0.69 0.36 3 0.18
e14
cision matrices. Take the subgroup g2 as an example, the adjusted in­
g5 e15, e18, 0.41 0.30 5 0.20 dividual decision matrices are
e19, e22, ⎛ ⎞
e24 0.90 0.70 0.32 0.80 0.20
⎜ 0.20 1.00 0.80 0.70 0.20 ⎟
R7 = ⎜ ⎟ 17
⎝ 1.00 0.10 0.30 1.00 0.30 ⎠, R
⎛ ⎞ 0.54 0.20 0.30 0.10 1.00
0.32 0.25 0.85 0.80 0.83 ⎛ ⎞
⎜ 0.30 0.82 0.35 0.25 0.25 ⎟ 0.90 0.60 0.30 0.80 0.40
1 ⎜ ⎟ 2
R =⎜ ⎟, R ⎜ 0.40 0.70 0.70 0.80 0.10 ⎟
⎝ 0.93 0.80 0.30 0.83 0.27 ⎠ =⎜ ⎝ 1.00 0.10 0.40 1.00 0.40 ⎠

0.68 0.35 0.78 0.13 0.87 0.45 0.10 0.20 0.10 0.60
⎛ ⎞
0.85 0.68 0.28 0.80 0.38
⎜ 0.23 ⎛ ⎞
⎜ 0.88 0.85 0.85 0.20 ⎟
⎟ 0.90 0.60 0.40 1.00 0.30
=⎜ ⎟ ⎜ 0.20
⎝ 1.00 0.23 0.23 0.93 0.38 ⎠ 0.90 1.00 0.90 0.10 ⎟
R21 = ⎜
⎝ 1.00
⎟, R25
0.68 0.13 0.30 0.10 0.90 0.60 0.10 0.90 0.70 ⎠
0.71 0.10 0.10 0.10 1.00
⎛ ⎞ ⎛ ⎞
0.26 0.76 0.79 0.31 0.83 0.70 0.80 0.20 0.60 0.60
⎜ 0.74 ⎜ 0.10 0.90 0.90 1.00 0.40 ⎟
⎜ 0.31 0.36 0.34 0.67 ⎟
⎟ 4 =⎜⎝ 1.00

R3 = ⎜ ⎟, R 0.10 0.10 0.80 0.10 ⎠
⎝ 0.34 0.27 0.74 0.87 0.81 ⎠ 0.43 0.10 0.60 0.10 1.00
0.16 0.70 0.31 0.70 0.36
⎛ ⎞
0.17 0.13 0.23 0.17 1.00 where the fairness index of subgroup g2 is Gi (gp) = 0.
⎜ 0.87
⎜ 0.27 0.27 0.90 0.10 ⎟
⎟ Meanwhile, the final group decision matrix is derived as follows:
=⎜ ⎟ ⎛ ⎞
⎝ 1.00 0.87 0.87 0.40 0.97 ⎠ 0.38 0.30 0.70 0.40 0.73
0.10 0.67 0.93 0.13 0.40 ⎜ 0.60 0.50 0.50 0.44 0.30 ⎟
Rc = ⎜
⎝ 0.96 0.40 0.64 0.80 0.75 ⎠

⎛ ⎞
0.74 0.20 0.80 0.30 0.26 0.22 0.35 0.66 0.14 0.70
⎜ 0.72 0.34 0.82 0.38 0.64 ⎟
R5 = ⎜
⎝ 0.96
⎟ Step 7: Based on Rc , the assessment values of alternatives are
0.28 0.78 0.36 0.80 ⎠
0.12 0.14 0.78 0.18 0.86 A (x1) = 0.49, A (x2) = 0.48, A (x3) = 0.73, A (x4) = 0.37.
Therefore, the ranking of the alternatives is x3 ≻ x1 ≻ x2 ≻ x4 , and
Moreover, the group decision matrix is derived as follows: X3 is the best choice.
⎛ ⎞
0.46 0.41 0.61 0.47 0.67
c
⎜ 0.57 0.52
⎜ 0.52 0.52 0.39 ⎟⎟ 6. Comparative analysis
R =⎝
0.83 0.48 0.59 0.69 0.64 ⎠
0.34 0.41 0.61 0.27 0.67 This section conducts sensitive and comparative analyses to further
validate the new method’s effectiveness.
By Eq. (3), the subgroup consensus levels are CI(g1) = 0.78, CI(g2) =
0.71, CI(g3) = 0.76, CI(g4) = 0.71, and CI(g5) = 0.79, respectively.
Meanwhile, the group consensus level is GCI = 0.75 according to Eq. 6.1. Sensitivity analysis
(10). Thus, the consensus adjustment should be activated.
Step 3: As CI(g4) = min{CI(gp) |p = 1, 2, …, 5}, and the size of When different values of the Gini coefficient and different consensus
subgroup g4 is 3, which is smaller than {⌊25 /5⌋, ⌊25 × 0.15⌋}, g4 is thresholds are used in the case study in Section 5, the corresponding
identified as a minority opinion subgroup. Then, DMs in subgroup g4 decision-making results are shown in Table 4.
elaborate on the rationality of its opinions. After a full discussion, sub­ Based on the results in Table 4, we derive the following conclusions:
groups g1, g2, and g3 support the opinions of g4, while subgroup g5 op­
∑ ∑
poses it. In this situation, we have p∈indGS,p∗ λp − p∈indGO,p ∗ λp = 0.62 − (i) For the fixed Gini coefficient, the higher the consensus threshold
0.20 = 0.42 > 0. It shows that the subgroup g4 is worth considering. is, the more the total opinion adjustments will be.
According to Eq. (11), the updated weight of subgroup g4 is (ii) For the consensus thresholds 0.80 and 0.85, the total opinion
adjustments increase with the decrease of the Gini coefficient.
λ4 = max{min{0.42, 0.21}, 0.18} = 0.21
However, this conclusion is contrary to the consensus threshold
Meanwhile, by Eq. (12), the updated weighting vector on the sub­ 0.90, namely, the total opinion adjustments increase with the
group set is λp = (0.20, 0.17, 0.22, 0.20, 0.19). increase of the Gini coefficient.
Step 4: Based on model (18), the consensus adjustments of subgroups (iii) Interestingly, x3 is always the best choice, and x4 is always the
worst.

10
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

Table 4
Decision-making results for different Gini coefficients and consensus thresholds.
The range of Gini coefficient θ = 0.80 θ = 0.85

Gi TOA RA Gi TOA RA

[0, 0.19] 0.19 3.65 x3 ≻ x1 ≻ x2 ≻ x4 0.19 7.80 x3 ≻ x1 ≻ x2 ≻ x4


[0.2, 0.29] 0.29 3.64 x3 ≻ x1 ≻ x2 ≻ x4 0.29 7.78 x3 ≻ x1 ≻ x2 ≻ x4
[0.3, 0.39] 0.39 3.63 x3 ≻ x1 ≻ x2 ≻ x4 0.39 7.77 x3 ≻ x1 ≻ x2 ≻ x4
[0.4, 0.6] 0.60 3.61 x3 ≻ x1 ≻ x2 ≻ x4 0.47 7.76 x3 ≻ x1 ≻ x2 ≻ x4
The range of Gini coefficient θ = 0.90 θ = 0.95
Gi TOA RA Gi TOA RA
[0, 0.19] 0.19 11.98 x3 ≻ x1 ≻ x2 ≻ x4 0.19 16.60 x3 ≻ x2 ≻ x1 ≻ x4
[0.2, 0.29] 0.24 11.97 x3 ≻ x2 ≻ x1 ≻ x4 0.22 16.60 x3 ≻ x2 ≻ x1 ≻ x4
[0.3, 0.39] 0.30 12.00 x3 ≻ x2 ≻ x1 ≻ x4 0.30 17.21 x3 ≻ x2 ≻ x1 ≻ x4
[0.4, 0.6] 0.40 12.06 x3 ≻ x1 ≻ x2 ≻ x4 – – –

Note: TOA: the total opinion adjustment of subgroups; RA: the ranking of alternatives. “-” indicates infeasibility because the maximum consensus level is 0.933 when
the Gini coefficient belongs to [0.4, 0.6]. Further, we make the sensitivity analysis from 0.80 because the initial consensus level is 0.752.

Fig. 2. Comparison of four consensus methods.

Table 5
Comparisons of different consensus methods for the consensus threshold 0.85
Method Consensus mechanism Fairness Initial GCL Final GCL Iteration Opinion adjustment amount Ranking

Method in (Tang et al., 2020) FI-CM √ 0.75 0.85 24 50.84 x3 ≻ x1 ≻ x4 ≻ x2


Method in (Dong et al., 2010) OM-CM × 0.74 0.85 1 51.64 x3 ≻ x1 ≻ x2 ≻ x4
Method in (Liao et al., 2022) OM–FI–CM × 0.67 0.85 22 46.43 x3 ≻ x2 ≻ x4 ≻ x1
The proposed method OM-CM √ 0.76 0.85 1 45.12 x3 ≻ x1 ≻ x2 ≻ x4

Note: FI-CM: the feedback iteration consensus mechanism; OM-CM: the optimization model consensus mechanism; OM–FI–CM: the optimization model based
feedback iteration consensus mechanism.

6.2. Comparisons methods in Tang et al. (2020) and Liao et al. (2022) requires 24 and 22
iterations to reach the consensus threshold, respectively, Fig. 2 only
To show the efficiency and merits of the new method, this subsection offers the results for the first four iterations.
compares it with three previous methods (Tang et al., 2020; Dong et al., To further show their differences, more compared aspects are shown
2010; Liao et al., 2022). Especially, Tang et al. (2020) proposed a in Table 5.
feedback mechanism for consensus reaching for LSGDM, which con­ Except for the group consensus threshold, all parameters are iden­
siders the fair distribution of changes in DMs. Based on the ordered tical to those in the corresponding methods. For example, in Tang et al.’s
weighted averaging (OWA) operator, Dong et al. (2010) proposed the method (Tang et al., 2020), the parameters b and ε in the improved
minimum adjustment consensus model for traditional GDM. Liao et al. K-means clustering method are 1 and 0.001, respectively. The
(2022) constructed an optimal model-based interactive consensus intra-subgroup consensus threshold is 0.9, the group consensus
reaching method for LSGDM regarding the minimum cost consensus threshold is 0.85, the convex combination parameter δ of intra-subgroup
model. and group consensus thresholds is 0.7, and the consensus iteration co­
To facilitate the simulation analysis, let the group consensus level in efficient τ is 0.9; in Liao et al.’s method (Liao et al., 2022), the deviation
Dong et al.’s model (Dong et al., 2010) be the average of all DMs’ threshold between individual adjusted judgments and corresponding
consensus levels, while each DM is a core member in Tang and Liao’s collective one is 0.2, and the consensus iteration coefficient τ is 0.9; in
model (Tang et al., 2020). The main results are shown in Fig. 2. Since Dong et al.’s method (Dong et al., 2010), the values of a and b in the

11
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

Table 6
Comparisons of different consensus methods for the case study in Liao et al. (2022).
Method Consensus mechanism Fairness Initial GCL Final GCL Iteration Opinion adjustment amount Ranking

Method in (Tang et al., 2020) FI-CM √ 0.83 0.90 34 16.62 x2 ≻ x1 ≻ x4 ≻ x3


Method in (Dong et al., 2010) OM-CM × 0.85 0.90 1 14.26 x2 ≻ x4 ≻ x1 ≻ x3
Method in (Liao et al., 2022) OM–FI–CM × 0.78 0.90 25 10.48 x2 ≻ x4 ≻ x1 ≻ x3
The proposed method OM-CM √ 0.85 0.90 1 9.60 x2 ≻ x4 ≻ x1 ≻ x3

ordered weighted average operator are 0.2 and 0.8. ➀ It fully considers the importance of the minority opinion subgroup,
For reproducibility, we next compare the new method with the other and the associated weight-updating measures are proposed accord­
three in the case of Liao et al. (2022), which is about the rationality of ing to the reconsidering results;
using four types of drugs in China. Since the initial consensus level of ➁ The consensus adjustment mechanism regards three aspects: the
subgroup fuzzy preference relations is 0.854, we here let the fairness of consensus adjustment of sub-groups/DMs, the adjustment
intra-subgroup consensus threshold be 0.90 in Liao et al.’s method (Tang range of the minority opinion subgroups, and the total consensus
et al., 2020), and the group consensus threshold is 0.9. The distance adjustment;
threshold is defined as 1 for obtaining five subgroups in the improved ➂ Under the fixed consensus adjustment, individual judgments are
density peak clustering. All the other parameters are the same as shown adjusted towards the group as closely as possible that ensure
above. The main results are provided in Table 6, similar to the above consensus-reaching efficiency.
case study.
According to Tables 5 and 6, the following conclusions are obtained: However, there are also some limitations of the new method:

(i) Three types of consensus mechanisms are adopted. Especially, ➀ It does not consider the non-cooperative behaviours of DMs, which is
Tang et al.’s FI-CM (Tang et al., 2020) mainly contains two steps: a crucial factor in consensus building;
➀ identifying subgroup and DM total adjustments sequentially, ➁ ➁ It restricts the real-valued judgments, which is unsuitable for
adjusting judgment with the lowest consensus level of identified decision-making with fuzzy judgment information;
DM, where the adjustment direction is the group judgment. Dong ➂ It is offered for multi-criteria decision matrices rather than prefer­
et al.’s method (Dong et al., 2010) employs the minimum total ence relations, another important representation of judgments.
adjustment-based quadratic programming model, which aggre­
gates individual opinions by the ordered weighted average 7. Conclusions
operator. Liao et al.’s method (Liao et al., 2022) uses the mini­
mum cost consensus-based model to determine the minimum The rational decision-making results should have a high consensus
total adjustment under pre-set distance and consensus thresholds. degree and a high degree of fairness. Considering these two aspects, we
It then sets the subgroup judgment as the adjustment direction of developed a new two-stage consensus model based on the Gini coeffi­
identified DMs’ judgments. cient for LSGDM. First, the density peak clustering method is improved
(ii) Methods in Tang et al. (2020) and Dong et al. (2010) disregard for dividing DMs into subgroups, and the weights of subgroups are
the fairness of consensus adjustment, and Tang and Liao’s determined by the silhouette coefficient, subgroup opinion deviation,
consensus adjustment mechanism (Tang et al., 2020) takes DMs’ and the cardinality of subgroups. Then, a new method for identifying the
consensus level as an adjustment basis. However, Tang et al.’s minority opinion subgroup is provided, and a new weight renewing
consensus adjustment mechanism cannot realize the fairness of method is raised, by which the weight of minority opinions with
consensus adjustment as it also embraces the lowest consensus important values is increased. In contrast, the weights of remaining
level priority principle. Especially, it cannot ensure the minimum minority opinions are unchanged or assigned to zero. Subsequently, the
(cost) adjustment as it adopts the feedback interaction mecha­ Gini coefficient is introduced to define the fairness of consensus
nism. The total adjustment of Tang et al.’s FI-CM (Tang et al., adjustment. Following these results, several consensus adjustment
2020) is smaller than Dong et al.’s OM-CM (Dong et al., 2010) models are built that fully consider the minority opinion and the
and Liao et al.’s OM–FI–CM (Liao et al., 2022). For the consensus adjustment fairness. From the offered example and comparison, the new
reaching efficiency, Dong et al.’s OM-CM (Dong et al., 2010) is method shows some good characteristics.
higher than Tang et al.’s FI-CM (Tang et al., 2020) and Liao In addition to the application to the case study in Section 5, we can
et al.’s OM–FI–CM (Liao et al., 2022). also apply it to other decision problems, such as supply chain manage­
(iii) Table 5 and 6 indicate that different methods may obtain various ment (Wang et al., 2018), site selection (Liu et al., 2017), job-shop
rankings. It is because their concrete formulations and research scheduling (Gao et al., 2020; Wang et al., 2022), COVID-19 vaccine
views are different. In practice, we suggest the DMs employ (Li and Liao, 2023) and resources management (Xiao et al., 2020). Be­
methods with more rationalities or adopt several methods to sides the minimum-consensus adjustment based decision making
make final decisions to avoid irrational results. methods, we can similarly study minimum cost-based consensus
(iv) Tables 5 and 6 show that the new method has the smallest total mechanism (Gong et al., 2015). Furthermore, the new method can be
adjustments and derives the same best objects as the three extended to more complex LSGDM problems, including LSGDM with
compared methods. Moreover, because the new method empha­ fuzzy decision-making information, missing decision-making informa­
sizes the fairness of consensus adjustment, the non-cooperative tion, or heterogeneous decision-making information. Moreover, with the
behaviours of DMs can be reduced. Meanwhile, the adjustment development of social media, the relationship between DMs has an
restriction of minority opinions ensures democracy and objec­ important influence on decision-making results. Therefore, paying
tivity of decision-making results. attention to LSGDM problems in social networks (Liang et al., 2022; Liu
et al., 2023b) is necessary. Notably, intelligent optimization algorithm is
Compared with previous research, the new method shows the an effective way to derive the decision results in LSGDM problems, such
following merits: as the genetic algorithm (Gholizadeh et al., 2021), evolutionary
multi-objective algorithm (Zhao and Zhang, 2020), adaptive polyploid
memetic algorithm (Dulebenets, 2021), metaheuristic algorithm (Pasha

12
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

et al., 2022), hyperheuristic algorithm (Singh and Pillay, 2022), diffused Liang, Y., Ju, Y., Dong, P., Zeng, X., Martínez, L., Dong, J., Wang, A., 2023. A sentiment
analysis-based two-stage consensus model of large-scale group with core-periphery
memetic optimizer algorithm (Dulebenets, 2023), on-dominant sorting
structure. Inf. Sci. 622, 808–841.
genetic algorithm (Yi et al., 2020), distributed guidance anti-flocking Liang, X., Guo, J., Liu, P., 2022. A large-scale group decision-making model with no
algorithm (Wang et al., 2021b), and improved non-dominated sorting consensus threshold based on social network analysis. Inf. Sci. 612, 361–383.
genetic algorithm III (Zhang and Wang, 2023). Li, X., Liao, H., 2023. A consensus model for large-scale group decision making based on
empathetic network analysis and its application in strategical selection of COVID-19
vaccines. J. Oper. Res. Soc. 74, 604–621.
CRediT authorship contribution statement Li, Y., Kou, G., Li, G., Peng, Y., 2022a. Consensus reaching process in large-scale group
decision making based on bounded confidence and social network. Eur. J. Oper. Res.
303, 790–802.
Fanyong Meng: Methodology, Formal analysis, Investigation, Li, Y., Ji, Y., Qu, S., 2022b. Consensus building for uncertain large-scale group decision-
Writing – original draft, Writing – review & editing, Supervision. Den­ making based on the clustering algorithm and robust discrete optimization. Group
gyu Zhao: Formal analysis, Methodology, Investigation, Writing – Decis. Negot. 31, 453–489.
Li, C., Dong, Y., Herrera, F., 2019. A consensus model for large-scale linguistic group
original draft, Writing – review & editing. Xumin Zhang: Methodology, decision making with a feedback recommendation based on clustered personalized
Formal analysis, Investigation, Writing – original draft. individual semantics and opposing consensus groups. IEEE Trans. Fuzzy Syst. 27,
221–233.
Liao, H., Wu, Z., Tang, M., Wan, Z., 2022. An interactive consensus reaching model with
Declaration of competing interest updated weights of clusters in large-scale group decision making. Eng. Appl. Artif.
Intell. 107. Art. no. 104532.
Liu, H., Li, Z., Zhang, J., You, X., 2018a. A large group decision making approach for
The authors declare that they have no conflict of interest. dependence assessment in human reliability analysis. Reliab. Eng. Syst. Saf. 176,
135–144.
Data availability Liu, Y., Fan, Z., You, T., Zhang, W., 2018b. Large group decision-making, LGDM) with
the participators from multiple subgroups of stakeholders: a method considering
both the collective evaluation and the fairness of the alternative. Comput. Ind. Eng.
No data was used for the research described in the article. 122, 262–272.
Liu, B., Shen, Y., Chen, X., Sun, H., Chen, Y., 2014. A complex multi-attribute large-group
PLS decision-making method in the interval-valued intuitionistic fuzzy environment.
Acknowledgement Appl. Math. Model. 38, 4512–4527.
Liu, B., Guo, S., Yan, K., Li, L., Wang, X., 2017. Double weight determination method for
This work is supported by the National Natural Science Foundation experts of complex multi-attribute large-group decision-making in interval-valued
intuitionistic fuzzy environment. J. Syst. Eng. Electron. 28, 88–96.
of China (No. 72371134), the Ministry of Education Humanities and Liu, P., Dong, X., Wang, P., 2023a. Limited budget-based consensus model for large
Social Science Foundation of China (No. 22YJ630061), the Natural group decision making with hesitant fuzzy linguistic information. Appl. Soft Comput.
Science Foundation of Changsha in China (No. kq2202112), and the 142. Art. no. 110368.
Liu, P., Li, Y., Wang, P., 2023b. Opinion dynamics and minimum adjustment-driven
Startup Foundation for Introducing Talent of NUIST (No. 2022r059). consensus model for multi-criteria large-scale group decision making under a novel
social trust propagation mechanism. IEEE Trans. Fuzzy Syst. 31, 307–321.
References Lorenz, M.O., 1905. Method of measuring the concentration of wealth. Publ. Am. Stat.
Assoc. 9, 209–219.
Lu, Y., Xu, Y., Herrera-Viedma, E., Han, Y., 2021. Consensus of large-scale group decision
Bazarova, N.N., Walther, J.B., McLeod, P.L., 2012. Minority influence in virtual groups: a
making in social network: the minimum cost model based on robust optimization.
comparison of four theories of minority influence. Commun. Res. 39, 295–316.
Inf. Sci. 547, 910–930.
Cui, T.H., Raju, J.S., Zhang, Z., 2007. Fairness and channel coordination. Manag. Sci. 53,
Ma, G., Li, X., Zheng, J., 2020. Efficiency and equity in regional coal de-capacity
1303–1314.
allocation in China: a multiple objective programming model based on Gini
Ding, R., Palomares, I., Wang, X., Yang, G., Liu, B., Dong, Y., Herrera-Viedma, E.,
coefficient and data envelopment analysis. Resour. Pol. 66. Art. no. 101621.
Herrera, F., 2020. Large-scale decision-making: characterization, taxonomy,
Mandell, M., 1991. Modelling effectiveness-equity trade-offs in public service delivery
challenges and future directions from an artificial intelligence and applications
systems. Manag. Sci. 37, 467–482.
perspective. Inf. Fusion 59, 84–102.
Meng, F., Tang, J., An, Q., 2023a. Cooperative game based two-stage consensus
Dong, Q., Cooper, O., 2016. A peer-to-peer dynamic adaptive consensus reaching model
adjustment mechanism for large-scale group decision making. Omega 117, 102842.
for the group AHP decision making. Eur. J. Oper. Res. 250, 521–530.
Meng, F., Chen, B., Tan, C., 2023b. Adaptive minimum adjustment consensus model for
Dong, Y., Xu, Y., Li, H., Feng, B., 2010. The OWA-based consensus operator under
large-scale group decision making under social networks and its application in
linguistic representation models using position indexes. Eur. J. Oper. Res. 203,
Integrated Care of Older People. Appl. Soft Comput. 132. Art. no. 109863.
455–463.
Meng, F., Zhao, D., Tan, C., Li, Z., 2023c. Ordinal-cardinal consensus analysis for large-
Dreu, C.K.D., West, M.A., 2001. Minority dissent and team innovation: the importance of
scale group decision making with uncertain self-confidence. Inf. Fusion 93, 344–362.
participation in decision making. J. Appl. Psychol. 86, 1191–1201.
Meng, F., Gong, Z., Pedrycz, W., Chu, J., 2023d. Selfish-dilemma consensus analysis for
Dulebenets, M.A., 2021. An Adaptive Polyploid Memetic algorithm for scheduling trucks
group decision making in the perspective of cooperative game theory. Eur. J. Oper.
at a cross-docking terminal. Inf. Sci. 565, 390–421.
Res. 308, 290–305.
Dulebenets, M.A., 2023. A diffused memetic optimizer for reactive berth allocation and
Meng, F., Chen, B., Wang, Z., 2022. A trust-based large-scale group decision making
scheduling at marine container terminals in response to disruptions. Swarm Evol.
consensus reaching framework for intercity railway public-private partnership
Comput. 80, 101334.
model selection. Neural Comput. Appl. 34, 19091–19115.
Du, Z., Yu, S., Xu, X., 2020. Managing noncooperative behaviors in large-scale group
Nie, R., Tian, Z., Wang, J., Luo, H., 2020. An objective and interactive-information-based
decision-making: integration of independent and supervised consensus-reaching
feedback mechanism for the consensus-reaching process considering a non-support
models. Inf. Sci. 531, 119–138.
degree for minority opinions. Expert Syst. 37, e12543.
Gao, D., Wang, G., Pedrycz, W., 2020. Solving fuzzy job-shop scheduling problem using
Pasha, J., Nwodu, A.L., Fathollahi Fard, A.M., Tian, G., Li, Z., Wang, H., Dulebenets, M.
DE algorithm improved by a selection mechanism. IEEE Trans. Ind. Inf. 28,
A., 2022. Exact and metaheuristic algorithms for the vehicle routing problem with a
3265–3275.
factory-in-a-box in multi-objective settings. Adv. Eng. Inf. 52, 101623.
Gini, C., 1914. Sulla Misura della concentrazione e della variabilità dei caratteri.
Qu, S., Li, Y., Ji, Y., 2021. The mixed integer robust maximum expert consensus models
Premiate officine grafiche C. Ferrari 73, 1203–1248.
for large-scale GDM under uncertainty circumstances. Appl. Soft Comput. 107. Art.
Gholizadeh, H., Fazlollahtabar, H., Fathollahi-Fard, A., Dulebenets, M.A., 2021.
no. 107369.
Preventive maintenance for the flexible flowshop scheduling under uncertainty: a
Ren, R., Tang, M., Liao, H., 2020. Managing minority opinions in micro-grid planning by
waste-to-energy system. Environ. Sci. Pollut. Res. https://doi.org/10.1007/s11356-
a social network analysis-based large scale group decision making method with
021-16234-x.
hesitant fuzzy linguistic information. Knowl. Base Syst. 189. Art. no. 105060.
Gong, Z., Zhang, H., Forrest, J., Li, L., Xu, X., 2015. Two consensus models based on the
Rodriguez, A., Liao, A., 2014. Clustering by fast search and find of density peaks. Science
minimum cost and maximum return regarding either all individuals or one
344, 1492–1496.
individual. Eur. J. Oper. Res. 240, 183–192.
Rousseeuw, P.J., 1987. Silhouettes: a graphical aid to the interpretation and validation of
Gou, X., Xu, Z., Herrera, F., 2018. Consensus reaching process for large-scale group
cluster analysis. J. Comput. Appl. Math. 20, 53–65.
decision making with double hierarchy hesitant fuzzy linguistic preference relations.
Sueyoshi, T., Qu, J., Li, A., Liu, X., 2021. A new approach for evaluating technology
Knowl. Base Syst. 157, 20–33.
inequality and diffusion barriers: the concept of efficiency Gini coefficient and its
Gou, X., Xu, Z., 2021. Managing noncooperative behaviors in large-scale group decision-
application in Chinese provinces. Energy 235. Art. no. 121256.
making with linguistic preference orderings: the application in Internet venture
Singh, E., Pillay, N., 2022. A study of ant-based pheromone spaces for generation
capital. Inf. Fusion 69, 142–155.
constructive hyper-heuristics. Swarm Evol. Comput. 72, 101095.
Gou, X., Xu, Z., Liao, H., Herrera, F., 2021. Consensus model handling minority opinions
and noncooperative behaviors in large-scale group decision-making under double
hierarchy linguistic preference relations. IEEE Trans. Cybern. 51, 283–296.

13
F. Meng et al. Engineering Applications of Artificial Intelligence 126 (2023) 106962

Tang, M., Liao, H., 2021. Failure mode and effect analysis considering the fairness- Xiong, K., Dong, Y., Zha, Q., 2023. A large-scale consensus model to manage non-
oriented consensus of a large group with core-periphery structure. Reliab. Eng. Syst. cooperative behaviors in group decision making: a perspective based on historical
Saf. 215. Art. no. 107821. data. Expert Syst. Appl. 214. Art. no. 119163.
Tang, M., Liao, H., Xu, J., Dalia, S., Zheng, X., 2020. Adaptive consensus reaching process Xu, X., Du, Z., Chen, X., Cai, C., 2019. Confidence consensus-based model for large-scale
with hybrid strategies for large-scale group decision making. Eur. J. Oper. Res. 282, group decision making: a novel approach to managing non-cooperative behaviors.
957–971. Inf. Sci. 477, 410–427.
Tian, X., Li, W., Xu, Z., Kou, G., Nie, C., 2022. An endo-confidence-based consensus with Xu, X., Du, Z., Chen, X., 2015. Consensus model for multi-criteria large-group emergency
hierarchical clustering and automatic feedback in multi-attribute large-scale group decision making considering non-cooperative behaviors and minority opinions.
decision-making. Inf. Sci. 608, 1702–1730. Decis. Support Syst. 79, 150–160.
Wan, S., Yan, J., Dong, J., 2022. Personalized individual semantics based consensus Xu, Y., Wen, X., Zhang, W., 2018. A two-stage consensus method for large-scale multi-
reaching process for large-scale group decision making with probabilistic linguistic attribute group decision making with an application to earthquake shelter selection.
preference relations and application to COVID-19 surveillance. Expert Syst. Appl. Comput. Ind. Eng. 116, 113–129.
191. Art. no. 116328. Yi, J., Xing, L., Wang, G., Dong, J., Vasilakos, A.V., Alavi, A.H., Wang, Ling, 2020.
Wan, S., Rao, T., Dong, J., 2023. Time-series based multi-criteria large-scale group Behavior of crossover operators in NSGA-III for large-scale optimization problems.
decision making with intuitionistic fuzzy information and application to multi- Inf. Sci. 509, 470–487.
period battery supplier selection. Expert Syst. Appl. 232. Art. no. 120749. Yu, S., Du, Z., Zhang, X., Luo, H., Lin, X., 2022. Trust cop-Kmeans clustering analysis and
Wang, G., Gao, D., Pedrycz, W., 2022. Solving multiobjective fuzzy job-shop scheduling minimum-cost consensus model considering voluntary trust loss in social network
problem by a hybrid adaptive differential evolution algorithm. IEEE Trans. Ind. Inf. large-scale decision-making. IEEE Trans. Fuzzy Syst. 30, 2634–2648.
18, 8519–8528. Zhang, S., Zhang, S., Ma, Z., Liu, X., 2023. A two-stage multi-attribute group consensus
Wang, Z., Rodríguez, R., Wang, Y., Martínez, L., 2021a. A two-stage minimum model based on distributed linguistic assessment information from the perspective of
adjustment consensus model for large scale decision making based on reliability fairness concern. Adapt. Behav. 31, 213–238.
modeled by two-dimension 2-tuple linguistic information. Comput. Ind. Eng. 151. Zhang, H., Wang, G., 2023. Improved NSGA-III using transfer learning and centroid
Art. no. 106973. distance for dynamic multi-objective optimization. Complex Intell. Syst. 9,
Wang, G., Wei, C., Wang, Y., Pedrycz, W., 2021b. Improving distributed anti-flocking 1143–1164.
algorithm for dynamic coverage of mobile wireless networks with obstacle Zhang, X., Meng, F., 2022. A large-scale group decision making method to select the ideal
avoidance. Knowl. Base Syst. 225. Art. no. 107133. mobile health application for the hospital. Appl. Intell. 52, 15844–15864.
Wang, P., Xu, X., Huang, S., Cai, C., 2018. A linguistic large group decision making Zhao, H., Zhang, C., 2020. An online-learning-based evolutionary many-objective
method based on the cloud model. IEEE Trans. Fuzzy Syst. 26, 3314–3326. algorithm. Inf. Sci. 509, 1–21.
Xiao, F., Chen, Z., Wang, X., Hou, W., Wang, J., 2022. Managing minority opinions in risk Zhong, X., Xu, X., Pan, B., 2022. A non-threshold consensus model based on the
evaluation by a delegation mechanism-based large-scale group decision-making with minimum cost and maximum consensus-increasing for multi-attribute large group
overlapping communities. J. Oper. Res. Soc. 73, 2338–2357. decision-making. Inf. Fusion 77, 90–106.
Xiao, J., Wang, X., Zhang, H., 2020. Managing personalized individual semantics and Zhong, X., Xu, X., Yin, X., 2021. A multi-stage hybrid consensus reaching model for
consensus in linguistic distribution large-scale group decision making. Inf. Fusion 53, multi-attribute large group decision-making: integrating cardinal consensus and
20–34. ordinal consensus. Comput. Ind. Eng. 158. Art. no. 107443.

14

View publication stats

You might also like