Expert Systems With Applications 237 (2024) 121560

Contents lists available at ScienceDirect

Expert Systems With Applications

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

A consensus method in social network large-scale group decision making

with interval information
Jiangjing Tan a, Yingming Wang a, b, *, Junfeng Chu a, b, *
a
School of Economics and Management, Fuzhou University, Fuzhou 350108, China
b
Decision Sciences Institute, Fuzhou University, Fuzhou 350108, China

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

Keywords: Large-scale group decision making (LSGDM) is considered when the number of experts involved in the decision
Large-scale group decision making exceeds 20. Due to the large number of people involved in LSGDM, its decision-making process is uncertain,
CRP complex, and time-consuming. Therefore, how to effectively help large groups reach consensus in a complex
Interval number
environment is a challenge for current research. Consensus reaching process (CRP) is an effective tool to elim­
Louvain algorithm
Social network
inate group conflicts. Based on this, we propose a consensus reaching process based on the Louvain algorithm,
social network, and bounded confidence (SNBC) model with interval numbers. First, we use interval numbers to
express expert opinions and social network relationships among experts. Second, the experts are clustered using
the Louvain algorithm. The weights of experts are obtained by social network analysis. Third, we use the SNBC
model to design a feedback mechanism for tripartite opinions. In addition, we give a numerical example and
simulation experiments to demonstrate the flexibility and effectiveness of the proposed approach. Finally, the
comparative analysis shows the superiority of our method.

1. Introduction based on uncertain data is common and sometimes unavoidable in many

real-world applications. Wu and Liu (2023) provided a global forum for
With the development of information technology and the complexity advancing the analysis, understanding, development, and practice of
of decision-making problems, more and more experts are willing to uncertainty theory and operations research for solving economic, engi­
participate in large-scale group decision making (Pelissari et al., 2020; neering, management, and social problems. A review paper provided a
Qin, Li & Liang, 2022; Zhang et al., 2021; Zhong, Xu & Pan, 2022). If comprehensive overview of fuzzy and linguistic decision making,
there are more than 20 experts, we define group decision-making as a enhancing the various extensions, applications, and challenges of the
large-scale group decision making (LSGDM) (Chen, Zhang, Xu & Cao, past 50 years (Herrera-Viedma et al., 2021). It is necessary to study how
2022a; Ding et al., 2020; Dong et al., 2018; Du, Luo, Lin & Yu, 2020). to help experts in LSGDM reach a consensus with a fuzzy language
LSGDM is increasingly practical in decision making environments and is environment. The second is cluster analysis approaches for large-scale
a research trend in decision science (Zhong, Xu & Yin, 2021). In LSGDM experts. To effectively reduce the difficulty of information processing,
problems, the experts with different levels of knowledge and back­ cluster analysis is often used to divide decision experts into subgroups.
grounds often express different opinions, which often prevents Many researchers studied the CRP model in LSGDM using different
consensus from being reached and leads to failed decisions (Herrera- clustering methods. Zhang, Dong and Herrera-Viedma (2018) proposed
Viedma et al., 2021). The consensus reaching process (CRP) is a key a new CRP model for the heterogeneous information in LSGDM and a
decision-making process for eliminating preference conflicts (Cheng selection process to divide the experts into several subgroups. Also, some
et al., 2022; Dong, Zha, Zhang & Herrera, 2021; Zhong et al., 2022). The researchers developed CRP method based on similarities or differences
current study of CRP in LSGDM focuses on three aspects. The first is the between experts’ opinions in LSGDM. For example, Wu, Liu, Qin and
language environment. In real life, it is difficult for experts to accurately Herrera (2019a) constructed a consensus matrix based on the similarity
express their preferences for alternative solutions with an accurate value of fuzzy preference relations, and then used the community detection
within a limited time (Deng, Xue & Jiang, 2023). Making decisions method to analyze the CRP. Chu, Wang, Liu and Liu (2020) presented a

* Corresponding authors at: School of Economics and Management, Fuzhou University, Fuzhou 350108, China.
E-mail addresses: [email protected] (J. Tan), [email protected] (Y. Wang), [email protected] (J. Chu).

https://doi.org/10.1016/j.eswa.2023.121560
Received 15 May 2023; Received in revised form 8 September 2023; Accepted 9 September 2023
Available online 12 September 2023
0957-4174/© 2023 Elsevier Ltd. All rights reserved.
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

two-stage CRP model for the sub-clusters to reach consensus in LSGDM 2021; Zhou et al., 2020; Zhou et al., 2022). A fuzzy number (Ma, Gong,
based on the fuzzy clustering method. Wu, Liu, Qin and Herrera (2022) Wei & Herrera-Viedma, 2022) is appropriate to reflect the complexity
proposed a new k-means clustering method that considers both prefer­ and uncertainty when experts are involved in complex environments.
ences and the preference adjustment cost. The shared feature in these There are various forms of representation have been proposed and
literatures is that experts are independent of each other. The third is the widely used in the literature, such as fuzzy sets (Herrera-Viedma,
feedback mechanism. Feedback mechanism is a key stage in improving Alonso, Chiclana & Herrera, 2007), intuitionistic fuzzy sets (Kumar &
consensus and reducing conflict in CRP (Liu, Xu & Herrera, 2019a). It Chen, 2022), interval number theory (Chen et al., 2022b). Second,
can be thought of as a dynamic process. If experts’ consensus level is existing clustering methods treat large groups as independent in­
below a given threshold, they are encouraged to negotiate until the dividuals. They usually calculate the opinion distance between experts
consensus level reaches the given threshold. and then group the experts according to the opinion distance. However,
With the development of social media, experts can transfer infor­ in reality, there are complex social network relationships between
mation and exchange opinions with each other, and others should in­ people. These approaches ignore the social network relationships of
fluence them who they trust in the social network. Some scholars use experts in large-scale decision-making. Third, the feedback mechanism.
social network analysis (SNA) to analyze group relationships in social The experts in the feedback mechanism based on the social network and
networks. SNA is a method to study the relationship between nodes in a bounded confidence tend to consider the opinions of experts whose
social network (Opsahl, Agneessens & Skvoretz, 2010). In the LSGDM opinion is less than a certain confidence value. However, experts only
literatures based on social network, the researchers used the relation­ refer to the opinions of these experts is relatively single, not compre­
ships between experts to promote the feedback mechanism of CRP in hensive and inefficient. In large-scale group decision making, it only
LSGDM. Tian, Nie and Wang (2019) built CRP in LSGDM, and completed refers to the opinions of a certain range of experts and ignores the
the social network based on SNA. Liu et al. (2019b) used trust propa­ opinions of the entire decision group. Since large-scale group decision
gation to complete the social network and then presented the CRP based making involves a large number of experts, but the reference objects of
on conflict detection and elimination in decision making to reach a experts’ opinions are limited and single, it is difficult to reach a
consensus. Li, Kou, Li and Wang (2021) proposed a multi-attribute group consensus in decision-making, which increases the time of decision
decision making with opinion dynamics to analyze the influence of the making. Motivated by the challenge to overcome the above limitations,
relationships between experts in CRP. Tan, Zhu, Cabrerizo and Herrera- this contribution is devoted to answer the two research questions: How
Viedma (2021) investigated a CRP considering dynamic trust in LSGDM to reasonably express experts’ decision-making preferences and social
in the traditional trust-based consensus model. Xu, Zhang and Chen network relationships? How to improve the efficiency of CRP in
(2020) proposed the CRP model in LSGDM for considering the experts’ LSGDM? In order to solve the above problems, we propose a consensus
relations in the social network and preference risks based on fuzzy set. In reaching process based on the Louvain algorithm and SNBC model of
most of the above-mentioned studies, group consensus was regulated LSGDM with interval number. The main contributions and novelties of
through social network relationships. This consensus feedback mecha­ this paper are obtained as follows:
nism might have overlooked the experts’ willingness to accept the
feedback recommendations. (1) Interval value is used to represent social network relationships
With the popularity of the opinion dynamics (Hassani et al., 2022; and preferences among experts. The social network relationship
Zhang, Dong, Zhang & Pedrycz, 2020), many scholars have begun to use between experts and the decision preferences among experts are
the opinion dynamics model to construct a feedback mechanism in interval number. Interval number theory can help deal with fuzzy
LSGDM. The bounded confidence model is one of the widely applied information. Interval number is often more consistent with peo­
models in opinion dynamics, reflecting individuals’ willingness to ple’s thinking habits.
accept advice from others (Hegselmann & Krause, 2002). Zha et al. (2) The Louvain algorithm is based on multi-level optimization
(2019) proposed a feedback mechanism for LSGDM based on the modularity. It’s simple, fast, accurate and efficient. It is consid­
bounded confidence to help experts reach a consensus. Zhang, Gao and ered to be one of the best performing community detection al­
Li (2020), by considering the leadership and bounded confidence levels gorithms. Therefor the Louvain algorithm based on the interval
of experts, devised a new feedback mechanism that could provide number social relationships between experts is applied to divide
acceptable advice to experts who needed to modify their opinions. Liu, the large-scale group into the small subgroups, which can
Li and Wang (2023) proposed a bounded confidence-based consensus improve decision efficiency.
optimization model and a new trust propagation method considering the (3) We propose a new feedback mechanism of consensus reaching
relative importance of trust degrees to generate adjustment opinions process based SNBC model. The feedback mechanism assumes
within the bounded confidence of subgroups. Li, Kou, Li and Peng that in addition to believing their own opinions, experts can refer
(2022) proposed a new feedback mechanism, which considered not only to the opinions of people they trust and the opinions of the group.
the trust relationships between experts, but also the bounded confidence
level of experts. LSGDM involves many experts, making information The structure of this paper is as follows: Definitions and notations,
transmission and opinion exchange complex. These consensus feedback SNA, and the Louvain algorithm, interval number are introduced in
mechanisms have advantages in handling LSGDM problems by consid­ Section 2. A new feedback mechanism of CRP in LSGDM is proposed in
ering experts’ social networks and their psychological behavior when Section 3. A numerical example is applied to verify the advantage of our
accepting advice. They not only provide a new way to solve the actual proposed approach in Section 4. The simulation experiment, comparison
LSGDM problem but also help improve decision efficiency and analyses and advantages of this method are conducted to illustrate the
credibility. feasibility of the proposed approach in Section 5. Section 6 concludes
Although the existing studies have made significant contributions to this paper.
the CRP of LSGDM, these methods still have some limitations. First, the
relationships between experts are mainly measured in terms of exact 2. Preliminaries
values. A crisp number (Urena, Chiclana, Melancon & Herrera-Viedma,
2019) is the simplest form to express the interaction. However, it is Here, we briefly introduce the basic concepts of LSGDM, social
difficult to use a precise value to measure the relationship between ex­ network, consensus reaching process, and interval number, which will
perts and expert opinions in real life. Therefore, it is necessary to t provide the theoretical basis for this research.
consider the uncertainty of the decision-making environment (Herrera-
Viedma et al., 2021; Liang, Labella, Wang & Rodríguez, 2023; Wu et al.,

2
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Fig. 1. The general framework of CRP.

2.1. Large-scale group decision-making ∑

d− (ei ) = aij 2
The group decision-making problem is a decision situation in which
j

several experts are required to decide on one or more alternatives as a

solution to a given problem (Kacprzyk, 1986). When the number of
Definition 5. . (Degree centrality). (Wasserman & Faust, 1994). Degree
experts is greater than 20, the problem becomes a LSGDM problem (Ding
centrality shows the sum of the expert ei directly connected to other experts. It
et al., 2020). The main elements of a typical LSGDM problem are as
can be defined as:
follows:

d(ei ) = d+ (ei ) + d− (ei ) 3

(1) A finite set of viable alternatives X = {x1 , x2 , ⋯, xm }(m⩾2).
(2) A set of experts E = {e1 , e2 , ⋯, eh } be a set of h experts in a group.
Usually, whenh ≥ 20, the group is considered as a large-scale
Definition 6.. (Closeness centrality). (Sabidussi, 1966). Closeness cen­
group. The experts’ weighting vector is W = (w1 , w2 , ⋯, wh )T ,
∑ trality depends on the sum of the distances from the expert ei to the other
where hp=1 WP = 1. experts. It can be defined as:
(3) A set of attributes C = {c1 , c2 , ⋯, cn } is a predefined set of attri­
butes and the weight of attribute as λ = {λ1 , λ2 , ⋯, λn }, where 1
∑n c(ei ) = ∑ 4
j=1 λj = 1. i=j dist(e i , ej )

(4) A(p) = (apij )m×n (p = 1, 2, ⋯, h) is the decision-making matrix pro­

where ei , ej ∈ V, and dist(ei , ej ) means the distance from the expert to
vided by the expert ep , where apij represents the alternative value
other experts, the smaller the sum, the shorter the path from this expert
for xi with respect to attribute cj .
to other experts and the closer this expert to other experts.

2.2. Social network 2.3. Consensus reaching process

Definition 1.. (Kay, 1977). A social network can be represented by the Definition 7.. (Li et al., 2022). Consensus reaching process (CRP) is a
graph G(E, V), where the set of nodes is V = {v1 ,v2 ,⋯,vn }, and the nodes of dynamic and interactive method used to reach a group decision. We use it for
edges are E = {(vl , vk )|vl , vk ∈ V; l ∕
= k}. In decision science, experts can be the purpose of obtaining a collective solution as close as possible to unanimous
represented by nodes and trust connections among experts can be represented agreement. CRP can eliminate and reduce the conflict among experts to an
by edges. acceptable degree for the group. The general process of CRP is depicted in
Definition 2.. (Kay, 1977). Let A = (alk )h×h be the adjacency matrix of Fig. 1. Before entering the CRP, experts make an assessment based on the
{
1, (vl , vk ) ∈ E decision problem, options, and attributes. The moderator first collects the
G(E, V), where alk = . Generally, the graph has two types of evaluation information. This evaluation information is then used to calculate
0, (vl , vk ) ∕
∈E
representations: 1) the undirected graph and 2) the directed graph. In the the consensus level. When the consensus level reaches a given threshold, it can
undirected graph alk = 1 implies that expert el trusts ek and expert ek trusts el . proceed directly to the selection process. Otherwise, the feedback adjustment
In the directed graph, the edge alk = 1 only implies that expert el trusts ek . is required.

Definition 3.. (In-degree). (Wasserman & Faust, 1994). In-degree shows

2.4. Interval number
the number of other experts that point to the expert ei . It can be defined as:
∑
d+ (ei ) = aki 1 Definition 8. (Chengzhong, 1989; You & Fan, 2002). A interval number
∼
k
is defined as a = [a− , a+ ](a− ⩽a+ ), where two elements a− ,a+ represent
∼
respectively the lower bound, upper bound of the interval number a. For two
∼ ∼
Definition 4.. (Out-degree). (Wasserman & Faust, 1994). Out-degree interval numbers a = [a− , a+ ] and b = [b− ,b+ ], some standard operators are
shows the number of the expert ei that point to other experts. It can be defined summarized as follows Eqs. (5)-(8). In this work, the lower bounds are
as: positive values such that a− ⩾0, b− ⩾0.

3
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Fig. 2. The framework of the method proposed in this paper.

∼ ∼ ∼ ∼
a + b = [a− + b− , a+ + b+ ], a − b = [a− − b− , a+ − b+ ] 5 We know that X = {x1 , x2 , ⋯, xm }(m⩾2) is a predefined set of alter­
natives, E = {e1 , e2 , ⋯, eh } is the experts in a group, W = (w1 , w2 , ⋯, wh )T
∑
is the experts’ weighting vector, where hp=1 WP = 1. C = {c1 , c2 , ⋯, cn }
∼
6
∼ ∼
a × b = [a− b− , a+ b+ ], k × a = [ka− , ka+ ]
is the set of attributes and the weight of attribute as λ = {λ1 , λ2 , ⋯, λn },
∼ ∼ ∼
7
∼ ∑ ∼
a/b = [a− /b+ , a+ /b− ], 1/a = [1/a− , 1/a+ ] where nj=1 λj = 1. Let A(p) = [apij ]m×n (p = 1, 2, ⋯, h) is the interval num­
[ −
In general, the distance between two interval numbers has the ber decision-making matrix provided for the expert ep , where ̃
p
aij = apij ,
following two methods: ]
The generalized distance formula is: ap+
ij indicates the alternative value for xi with respect to attribute cj .

∼ ∼
d(a, b) = max(|a− − b− |, |a+ − b+ |) 8 3. A CRP based on the Louvain algorithm and SNBC model
The Frobenius norm distance is:
⃦∼ ∼⃦ In this section, we present a consensus reaching process based on the
∼ ∼ ⃦ ⃦
d(a, b) = ⃦a − b⃦ = |a− − b− | + |a+ − b+ | 9 Louvain algorithm and SNBC model.
Our method of the CRP based on the Louvain algorithm and SNBC

4
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Fig. 3. The steps of the Louvain algorithm.

model is applied to LSGDM in social networks with interval number, meaning is the difference between the number of node’s sides in the com­
which includes three processes: the social network analysis process, the munity and the number of edges in a random case, and the equation is shown
consensus reaching process, the selection process. The general process of as:
the method in Fig. 2.
[(∑ ) (∑ )2 ]
∑
Q= c in
− tot
10
2M 2M
3.1. Social network analysis
∑ ∑
Where in represents all the weights within the community c, tot
This section uses the Louvain method to divide the large-group into represents the weights of the edges connected to the nodes inside the
several sub-clusters. And the degree centrality and closeness centrality community c, including the edges inside the community and the edges
are calculated by SNA to measure the weight of experts and sub-clusters. outside the community. M is the sum of all the edge weights in the
The Louvain method is a new community detection algorithm based on network.
modularity (Blondel, Guillaume, Lambiotte & Lefebvre, 2008; Waltman
& van Eck, 2013) and it has the following advantages: Definition 10.. (Blondel et al., 2008). Accordingly, the definition of the
incremental change of modularity is given as:
• High efficiency, suitable for large-scale networks. [∑ ] [∑ ]2 [∑ ] [∑ ]
• Automatic determination of community number. ΔQ =
+ 2ki,in
in
− tot + ki
− in
+ tot ki
11
• Widely applicable in complex networks. 2M 2M 2M 2M 2
• Flexibility with adjustable resolution parameter. ∑
Where in represents all the weights within the community c. tot
∑
• Modularity optimization, finding highly cohesive communities. represents the weights of the edges connected to the nodes inside the
community c, including the edges inside the community c and the edges
outside the community c. ki,in represents the total weights of the node vi
Definition 9.. (Waltman & van Eck, 2013). Modularity is an important in the community c. ki represents the total weight of node vi .
measure to evaluate the quality of a community network. Its physical The Louvain algorithm have the following properties.

5
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Property 1.. The Louvain algorithm maximizes modularity Q by

converging on the cumulative computation of all nodes in the community.
This model can be optimized through a strategy to continuously construct a
community structure with internal aggregation and external sparse connec­
tion. After an iteration, if there is no change, the process terminates; other­
wise, iteration continues until convergence.
Property 2.. If a community node is completely ‘‘closed’’, that is, all nodes
are internally connected to each other, but are not connected to other nodes
outside the community, then Q = 1 in Eq. (10). Fig. 4. A simple network.
Property 3.. The Louvain algorithm has two stages. In the first, the nodes
in the network are continuously traversed, and it is attempted to add to the in the community c. K−i.,in represents the total left interval weights of the
community a single node that can maximize ΔQ until all nodes no longer node vi in the community c. K+ i.,in represents the total right interval
change. In the second stage, edges from multiple nodes in the same community ∼
weights of the node vi in the community c. Ki = [K−i , K+ i ] represents the
to another community in the original network are represented by a weighted
total interval weight of the node vi . K−i represents the total left interval
edge in the new network. The steps of the Louvain algorithm in Fig. 3.
weight of the node vi . K+i represents the total right interval weight of the
In the previous Louvain algorithm, the weights between nodes are node vi .
mostly exact values. However, using interval number weights instead of Below is a brief introduction on how to calculate the change in
exact value weights to perform the Louvain algorithm is more reason­ modularity when dealing with interval-valued edge weights. Assuming
able. Therefore, extending the operation range of the Louvain algorithm we have a simple network with three nodes (See Fig. 4). Nodes 1 and 2
from exact value to interval number operation is necessary. Here, we are already in the same community, while node 3 is separate.
give the modularity formula of the Louvain algorithm for interval We want to calculate the modularity gain when node 3 joins the
numbers. community of nodes 1 and 2. Given weight intervals:Weight(1, 2) =
∑ ∼ ∼
Definition 11.. The modularity formula of the Louvain algorithm based [3, 5],Weight(1, 3) = [4, 6]. Here ∼ in = [3, 5], ki,in = [4, 6] , ki = [4, 6].
on interval number is defined as: First, calculate the modularity gain using the lower bounds of the
weights.
[(∑∼ ) (∑∼ )2 ]
∼ ∑ ∑ ∑ ∑
Q= c ∼
in
− tot
∼ (1) Calculate node weight: −tot1 = 3 + 4 = 7, −tot2 = 3, −tot3 = 4.
2M 2M (2) Calculate the total weight M of the network:M = (7 + 3 + 4)/
− −

[
∑
[(∑− ) (∑ ) ]
− 2 ∑
[(∑+ ) (∑ )2 ] ]
+ 12 2 = 7.
= c in
− − tot
, c in
+ − tot (3) Use the Eq. (13) for modularity gain, substituting in the values.
2M 2M − 2M 2M +
We get:
= [Q− , Q+ ] [ ] [ ]2 [ ] [ ]
3+2×4 7+3+4 3 (7 + 3) × 4
∑∼ ∑− ∑+ ΔQ− = − − + = − 0.0204
Where in = [ in , in ] is the sum of all the interval edge weights in 2×7 2×7 2×7 2×7 2
∑
the community c, −in is the sum of all the left interval edge weights in
∑+ Based on the lower bound weights calculated above, the modularity
the community c, in is the sum of all the right interval edge weights in
∑ ∑− ∑+ gain for the lower bound is − 0.0204.
the community. ∼ tot = [ tot , tot ] is the sum of all the interval edge
∑ Secondly, we can use a similar approach with the upper bound
weights related to all nodes in the community c. −tot is the sum of all the weights for comparable calculations. The resulting upper bound
∑
left interval edge weights related to all nodes in the community c. + tot is modularity gain is:
the sum of all the right interval edge weights related to all nodes in the [ ] [ ]2 [ ] [ ]
∼ 5+2×6 11 + 5 + 6 5 (11 + 5) × 6
community c. M = [M− , M+ ] is the sum of all the interval edge weights in ΔQ+ = − − +
the network. M− is the sum of all the left interval edge weights of the 2 × 11 2 × 11 2 × 11 2 × 112
network. M+ is the sum of all the right interval edge weights in the = − 0.0579
network. Q− is the modularity of left interval in the network, Q+ is the
Finally, combining both lower and upper bounds, the interval for
modularity of the right interval network structure.
modularity gain is [-0.0579, − 0.0204]. A negative modularity gain
Definition 12.. Modularity gain is achieved by removing a node from its implies that adding node 3 to the community of node 1 and node 2
group and moving it to a neighboring group. The modularity gain is shown in would result in a decrease in modularity. In practice, we would opt for
Eq. (13): communities that offer positive or maximum modularity gain to merge
nodes, in order to optimize community structures and maximize
⎡∑∼ ∼ ⎤ [∑
∼
∼ ]2
⎡∑∼ ∼ ⎤
[∑∼ ] modularity.
+ 2k i,in + ki ki
To sum up, the Louvain algorithm for interval number shows in the
∼
ΔQ = ⎣ ∼
in ⎦− ∼
tot
− + ⎣ ∼tot2 ⎦
∼
in

2M 2M 2M 2M Algorithm 1.
⎡ [∑ Algorithm 1.
− −
] [∑ ]2 [ − ] [ − − ] ⎤
∑ ∑
+ 2ki,in − k Input: The social network node with interval number weights.
−
⎢ tot + ki tot i
⎢
in
−
− − in
−
+ 2
,⎥⎥ Output: The obtained sub-clusters gγ (γ = 1, 2, ⋯, k) and modularity.
13
⎢ 2M 2M − 2M − ⎥
2M
⎢ ⎥ Step 1. Let each node in the graph as an independent community with the same
⎢ ⎡ ⎤ ⎥
= ⎢ [∑ ] [ ] ∑+ ] ∑+ ⎥ number of communities as the number of nodes.
⎢ + ∑ 2 [ ⎥
⎢ + 2k + +
+ k + k +
⎥ Step 2. For each node, try to assign a node to the community where each of its
⎢ in i,in tot i in ⎢ tot i ⎥ ⎥
− − + ⎣ ⎦ ∼
⎣ + 2M + + 2 ⎦ neighbors is located. Calculate the modularity gain ΔQ before and after allocation
2M 2M 2M + ∼ ∼
using Eq. (13). If the neighbor node with the max ΔQ and the max ΔQ > 0, then the
node is assigned to the community; otherwise, it remains unchanged.
= [ΔQ− , ΔQ+ ] Step 3. Repeat Step 2 until the community of all nodes no longer changes.
∼ (continued on next page)
Ki,in = [K−i.,in , K+
i.,in ] represents the total interval weights of the node vi

6
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

(continued ) (continued )
Algorithm 1. Algorithm 2.

Step 4. Compress all nodes in the same community into a new node. The weight of Step 4: Use SNA and the number of experts to obtain the weight of sub-cluster gγ (γ = 1,
edges between nodes in the community is transformed into the weight of the new 2, ⋯, k). Normalize the degree centrality and closeness centrality of sub-cluster
node, and the edge weights between communities are transformed into edge gγ (γ = 1, 2, ⋯, k) as follows:
∑nγ
weights between new nodes.
p=1 d(ep )
Step 5. Repeat Step 1 until the modularity of the entire graph no longer changes. ds (gγ ) = ∑ (∑n ) 17
k
p=1 d(ep )
γ

Step 6. When the community does not change, output the sub-clusters gγ (γ = 1, 2, ⋯, γ=1
∑nγ
k) and modularity.Step 7. End. p=1 c(ep )
cs (gγ ) = ∑ (∑n ) 18
k
p=1 c(ep )
γ
γ=1

Step 5: The number Nγ = (N1 , N2 , ⋯, NK )T of the sub-cluster gγ (γ = 1, 2, ⋯, k) can be

calculated as follows:
Based on Algorithm 1,h experts can be divided into k sub-clusters gγ (γ = N2 (gγ )
Nγ = ∑ 2 19
N (gγ )
1, 2, ⋯, k), the sub-clusters’ weighting vector is G = (G1 , G2 , ⋯, Gk )T ,
∑ Where N(gγ ) represent the number of experts in sub-cluster gγ (γ = 1, 2, ⋯, k). Clearly,
where kγ=1 Gγ = 1. The experts’ weighting vector in each sub-cluster ∑
0⩽Nγ ⩽1(γ = 1, 2, ⋯, k) and kγ=1 Nγ = 1.
( )T
gγ (γ = 1, 2, ⋯, k) is wγ = wγ1 , wγ2 , ⋯, wγnγ , where nγ is the number of Step 6: Combine the above three indexes, the weight of the sub-cluster gγ (γ = 1, 2, ⋯
∑γ ds (gγ ) + cs (gγ ) + Nγ
experts in sub-cluster gγ (γ = 1, 2, ⋯, k) and np=1 wγp = 1. , k) can be calculated as follows:Gγ =
3
20

In this study, we selected degree centrality, closeness centrality, and Step 7: End.
the number of experts as indicators for calculating weights. Below are
the reasons for choosing these indicators:
3.2. Consensus reaching process
• Degree Centrality: Degree centrality measures the number of con­
nections a node has with other nodes. Selecting degree centrality as After converting from large-scale to small sub-clusters, the consensus
one of the indicators stems from the idea that the number of con­ level of the expert, sub-clusters, and group can be obtained and evalu­
nections a node possesses might be correlated with its influence ated to judge whether there is a consensus between the sub-clusters. If
within the network. not, the sub-cluster whose consensus level is the lowest is identified. The
• Closeness Centrality: Closeness centrality quantifies the average feedback mechanism based on social network and the bounded confi­
shortest path length from a node to other nodes. We chose closeness dence model is used to help the experts in the sub-cluster whose
centrality because in certain scenarios, a node’s efficiency in infor­ consensus level is the lowest to adjust their opinions.
∼
mation dissemination and communication might impact its
∼
Let Ap = [ap,γ,t
ij ]m×n represents the decision-making matrix of the ep
γ,(t)
importance. ∼ ∼
• Number of Experts: In this study, we consider the count of experts as in sub-cluster gγ at the time t, the decision-making A(t)
γ = [aij ]m×n of the
γ,t

another weighting indicator, as having a larger number of experts sub-cluster gγ is Weighted Arithmetic Average of corresponding decision
might imply a higher level of knowledge or capability within a matrices of experts in the sub-cluster gγ at the time t:
specific domain.
∼ nγ
∑ ∼ nγ
∑ ∼
A(t)
γ = wγp Aγ,(t) = wγp ap,γ,t 21
In our methodology, we assume that these three factors are equally p=1
p
p=1
ij

important, and thus, we opt for an equal-weight combination. We start

by normalizing each indicator to ensure they lie within a comparable
∼ ∼
The final group decision-making matrix Aall = [atij ]m×n is the Weighted
(t)
numerical range. Subsequently, we sum the normalized scores of these ∼
indicators to obtain a composite score, representing the weight of each Arithmetic Average of sub-clusters matrices A(t)
γ at the time t:
entity. Given our assumption of equal importance, we utilize an average- ( n ) ( n )
weight approach for combining them. However, adjusting weight dis­ ∑ ∑ ∑ ∑ ∑
k ∼ k γ ∼ k γ ∼
(t)
Aall = (t)
Gγ Aγ = Gγ γ γ,(t)
w p Ap = Gγ γ p,γ,t
wp aij 22
tribution in different contexts might lead to distinct outcomes. For γ=1 γ=1 p=1 γ=1 p=1
instance, emphasizing information propagation could increase the
weight assigned to closeness centrality, while focusing on social influ­
∼ ∼
Let Aγ,(t)
p = [ap,γ,t
ij ]m×n be the decision matrix of the ep in sub-cluster gγ
ence might elevate the weight of degree centrality. In practical appli­ ∼ ∼
cations, this method can be adjusted based on different needs and and A(t)
γ = [aij ]m×n be the matrix of sub-cluster gγ at the time t, then
γ,t

contexts, enabling the modification of weights for more accurate ana­ ∼ ∼

d(Aγ,(t)
p , Aγ ) based on Manhattan distance can be calculated as follows:
(t)
lyses and decisions. The main steps for calculating the weight of experts
and sub-clusters are provided as follows: ∼ ∼
1 ∑ m ∑ n ⃒ ∼
⃒ γ,(t)
∼ ⃒
⃒
Algorithm 2. d(Aγ,(t) (t)
p , Aγ ) = ⃒Ap − A(t)
γ ⃒
m × n i=1 j=1
Input: The social network of experts. 23
n ⃒ ∼ ∼ ⃒
Output: The weight of experts and sub-clusters. 1 ∑ m ∑
⃒ p,γ,t γ,(t) ⃒
Step 1: Use SNA to obtain the degree centrality and closeness centrality Eqs (1)-(4) for
= ⃒aij − aij ⃒
m × n i=1 j=1
experts.
Step 2: Normalize the degree centrality and closeness centrality of the expert ep as
According to the Manhattan distance of the interval decision matrix,
follows:
d(ep )
experts’ consensus level, sub-clusters’ consensus level, and group’s
ds (ep ) = ∑nγ 14 consensus level are obtained.
p=1 d(ep )
c(ep ) Level 1: The consensus level of the expert ep at the time t can be
cs (ep ) = ∑nγ 15
p=1 c(ep ) defined as:
Step 3: Combine the above two indexes, the weight of the expert ep in the sub-cluster ∼ ∼ ∼
gγ (γ = 1, 2, ⋯, k) can be calculated as follows: ICI(Aγ,(t) γ,(t) (t)
24
P ) = 1 − d(AP , Aγ )
ds (ep ) + cs (ep )
wp =
γ
16
2 Level 2: The consensus level of the sub-cluster gγ at the time t can be
(continued on next column)
defined as:

7
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

nγ
∑ ∼ and the opinions of the whole group:
ICI (t) (gγ ) = wγp ICI(Aγ,(t)
p ) 25
p=1 ( ∼ )
(1 - θP ) ∑Aγ,(t)
∼ ∼ ∼
31
q (t)
Aγ,(t+1) = θP ⋅Aγ,(t) + ⋅ + Aall whereeq ∈ IP
Level 3: The group consensus level at the time t can be defined as: p p
2 q
NIp
( n )
∑k
( ) ∑
k ∑γ ( ) ∼
(t)
GCI = (t)
Gγ ⋅ICI gγ = Gγ γ
wp ICI Ap γ̃,(t)
26 Where Aγ,(t+1)
p represents the decision-making matrix of the ep in sub-
γ=1 γ=1 p=1 ∼
cluster gγ at the time t + 1. θp is the self-confidence value of the ep . Aγ,(t)
p
Experts usually have different industry backgrounds and knowledge
represents the decision-making matrix of the ep in sub-cluster gγ at the
levels. Therefore, a consensus threshold α needs to be determined for ∼
judging whether each subgroup reaches consensus. If GCI(t) ⩾α, it illus­ time t. Aγ,(t)
q represents the decision-making matrix of the eq in sub-
trates that all experts reach a consensus and experts can make a decision, cluster gγ at the time t, where eq ∈ IP .NIP is the number of experts in IP .
otherwise experts cannot reach a consensus, and they need to adjust ∼
Aall is the final group decision-making matrix at the time t.
(t)
their opinions by the feedback mechanism.
Bounded confidence means that two experts trust each other when IP = ∅, three cases are given as follows:
the distance between their opinions is less than or equal to a given Case a: TP ∕ = ∅, it shows that the experts trust other experts in the
threshold value. Next, the bounded confidence model of multi-attribute sub-cluster gγ , the expert ep can be influenced not only by his own
groups is as follows: opinion, but also the opinions of the expert with the highest consensus
level in set TP and the whole group:
n ⃒ ∼ ∼ ⃒
1 ∑ m ∑
∼ ∼
⃒ γ,(t) ⃒ ( ∼ )
d(Aγ,(t) γ,(t)
P , Aq ) = ⃒A − Aγ,(t)
q ⃒ ∼ ∼
(1 - θP ) ∼
m × n i=1 j=1 P Aγ,(t+1) = θP ⋅Aγ,(t) ⋅ argmaxICI(Aγ,(t) (t)
32
+ q ) + Aall
27
p p
2 eq ∈TP
n ⃒ ∼ ∼ ⃒
1 ∑ m ∑
⃒ p,γ,(t) ⃒
= ⃒aij − aq,γ(t)
ij ⃒ Case b: TP = ∅, but BP ∕ = ∅. This indicates that the expert ep does not
m × n i=1 j=1
believe any other expert in the sub-cluster gγ , but the distance to them is
Step 1: Identify the bounded confidence set of the expert ep : less than or equal to the bounded confidence value βp . Thus, experts ep
can be influenced not only by their own opinions, but also by the
{ ⃒ ∼ ∼ ⃒ }
⃒
= p) ∈ gγ ⃒d(Aγ,(t)
Bp = ep (q ∕ γ,(t) ⃒
28 opinions of the experts with the highest level of consensus in the set BP
P , Aq )⃒⩽βP
and by the opinions of the whole group:
Step 2: Calculate the influence of weight: ∼ ∼ ( ∼ ∼ )
(1 - θP )
⎧ Aγ,(t+1)
p = θ ⋅A
P p
γ,(t)
+ ⋅ argmax ICI(Aγ,(t)
q ) + A (t)
all 33
⎨ 1 , ep ∈ Bp
⎪ 2 eq ∈BP

wpq (t) = NBP p = 1, 2, ⋯, n.t = 0, 1, 2, ⋯ 29 Case c: TP = ∅ but BP = ∅. It shows that the expert ep distrust any
⎪
⎩
0, ep ∕
∈ Bp other experts in the sub-cluster gγ , and the distance with them is less than
Where NBP is the number of experts in BP . Clearly, wpq (t)⩾0 and or equal to the bounded confidence value βp . The expert ep can be
∑n influenced by the opinions of the whole group:
q=1 wpq (t)
=1.
Step 3: Opinion dynamics of the expert ep : ∼ ∼ ∼
34
(t)
Aγ,(t+1)
p = θP ⋅Aγ,(t)
p + (1 - θP )⋅Aall
∑ ∼
γ,(t)
Step 2: The CRP is a process of iteration, and the feedback mecha­
∼ ∼ ∼
eq ∈BP Aq
∼
Aγ,(t+1) = = wp1 Aγ,(t) + wp2 Aγ,(t) + ⋯ + wpn Aγ,(t)
n (i = 1, 2, ⋯, n)
nism is to improve the consensus level. Let t = 0, if ICI(t) < α, we should
P 1 2
N BP
30 identify the min{ICIt (gγ )}, then let t = t +1 and use the feedback mech­
LSGDM aims to obtain the most suitable alternative quickly and anism (Step 1) to adjust the experts’ opinions in the sub-cluster gγ which
effectively. When ICI(t) < α, the feedback mechanism should be con­ is min{ICIt (gγ )} and then assess whether the sub-clusters reached a
structed. The corresponding modification suggestions should be pro­ consensus. If ICI(t) ⩾α, it indicates that experts reach a consensus.
vided for experts in the sub-cluster with the lowest consensus level to Otherwise, let t = t +1 repeat Step 1 until ICI(t) ⩾α.
approach a consensus. Suppose that experts tend to accept the opinions
of experts they trust and the overall opinions. The experts tend to accept
advice if the distances between these opinions and their own opinions 3.3. Selection process
are less than or equal to a certain confidence value. At the same time,
experts also consider the opinions of the whole group. The feedback Once a predefined level of consensus is reached, a selection process is
mechanism based the social network and bounded confidence (SNBC) used to generate a ranking of alternatives from the final group matrix. To
model is constructed as follows: avoid confusion, the final group matrix with a consensus will still be
In the social network, when expert ep and eq are connected we can denoted as Aall = (Atij )m×n . The ranking of alternatives can be generated
(t)

denote as apq = 1. The set of experts connected with expert ep in the sub- based on the assessable value S(xi ) of alternative xi over other alterna­
cluster gγ can be represented by TP , where TP = {eq |apq = 1}. Give the tives, where the alternatives with higher values are ranked higher. In
bounded confidence value βp and the self-confidence value θp of the ep . this paper, applying the WA operation to fuse all the values in the ith row
The bounded confidence sets of the expert ep in the sub-cluster gγ :Bp = of Aall = (Atij )m×n , the assessable value S(xi ) of alternative xi is calculated
(t)

{ ⃒ ∼ ∼ ⃒ }
⃒ γ,(t) ⃒ as follows:
ep (q ∕
= p) ∈ gγ ⃒d(AP , Aq )⃒⩽βP . The set of experts trusted by expert
γ,(t)
( )
ep in the sub-cluster gγ can be represented by IP , where IP = {BP ∩ TP }. ∑
n ∑
n ∑
k
S(xi ) = λj A(t) = λj Gγ A(t)
Step 1: When the sub-cluster gγ with the lowest consensus level can j=1
all
j=1 γ=1
γ

be identified, then all the experts in the sub-cluster gγ need to adjust their ( ( )) 35
∑
n ∑
k nγ
∑
opinions. Take expert ep for example. There are two possible = λj Gγ wγp Aγ,(t)
p
scenarios:IP ∕
=∅ j=1 γ=1 p=1
IP ∕
= ∅, the expert ep can refer to the opinions of the experts in set IP

8
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Fig. 5. The social network of 25 experts.

Step 2: Dimensionality reduction processing.

Table 1
This section adopts the social network of Li and Wei (2020) to vali­
The obtained sub-cluster gγ .
date the effectiveness of the proposed approach with 25 experts, four
Sub-clusters Color Expert members alternatives, and four attributes. The social network and interval num­
g1 yellow {e1 , e2 , e3 , e4 , e5 , e6 , e7 , e8 , e9 } ber weight of relationships between experts are shown in Fig. 5. Using
g2 pink {e10 , e12 , e12 , e13 , e14 , e15 , e16 , e17 , e18 , e19 , e20 } the Louvain algorithm for interval numbers to reduce the dimension of
g3 green {e20 , e21 , e22 , e23 , e24 , e25 }
social network and to obtain three sub-clusters. The specific results are
shown in Table 1 and Fig. 6. The interval modularity is [0.616,0.632]
4. A numerical example (Algorithm 1).
Step3: Social network analysis.
In this section, a numerical example is given to illustrate the effec­ The degree centrality and closeness centrality of each expert can be
tiveness and the advantages of the proposed approach. obtained by SNA, then the weight wγp of the expert ep and the sub-clusters
Step 1: Problem description and definition. weight Gγ can be computed (Algorithm 2). The specific results are
In the early days of the COVID-19 outbreak, the hospital mainly used shown in.
manual paper registration to manage the information of people. The Table 2 and Eqs. (14)-(20).
average time for manual registration is about 5–6 min. There are many
G1 = 0.3429 , G2 = 0.4529, G3 = 0.2042
disadvantages in this method, including the risk of virus cross infection,
the low inspection efficiency, which bring great hidden trouble and risk Step 4: Collecting preference matrix and compute the consensus
to the management work. In response to the need for epidemic infor­ level.
mation management, the hospital plans to select a COVID-19 informa­ The interval number decision matrix of experts and the group pref­
tion management system. There are four hospital information systems to erence matrix for the four attributes of the systems show in and Table 3.
choose from: System 1 (x1 ), System 2(x2 ), System 3 (x3 ), and System 4 Let t = 0, and AP
γ,(t)
= (ap,γ,t
ij )4×4 (P = 1, 2, ⋯, 25). By Eqs. (21)-(26),
(x4 ). The following four attributes are evaluated: security (c1 ), accuracy we can obtain the sub-cluster consensus level, where ICI(0) (C1 ) =
(c2 ), high efficiency (c3 ), convenience(c4 ). The weight of the attributes
[0.7056,0.7063],ICI(0) (C2 ) = [0.7032,0.7042], ICI(0) (C3 ) =
λ = (0.3, 0.3, 0.2, 0.2)T . The 25 experts E = {e1 , e2 , ⋯, e25 } participated
[0.6732,0.6776], and GCI = [0.6986,0.6988]. As the consensus
(0)
in the selection of the hospital information system. Assume that the ( )〈
threshold α is 0.8 and min GCI(0) α, a feedback mechanism is used to
consensus threshold α is 0.8. Denote the self-confidence value and the
bounded confidence level of the 25 experts are. adjust experts’ decision matrix.
Step 5: Using the feedback mechanism to adjust the experts’ decision
θ= matrix.
( )T Round 1:
0.6, 0.8, 0.6, 0.6, 0.8, 0.5, 0.6, 0.7, 0.5, 0.5, 0.7, 0.6, 0.7, 0.8, 0.6, As minICIγ(0) (Cγ ) = ICI(0) (C3 ), to be sure, experts in the sub-cluster C3
and
0.5, 0.6, 0.5, 0.7, 0.6, 0.7, 0.6, 0.5, 0.8, 0.7
( )T need to revise their opinions according to the feedback mechanism in
0.4, 0.3, 0.1, 0.15, 0.4, 0.1, 0.35, 0.1, 0.3, 0.1, 0.4, 0.3, Section 3.2, where C3 = {e20 ,e21 ,e22 ,e23 ,e24 ,e25 }. For e20 , e21 as example:
β = .
0.35, 0.4, 0.3, 0.3, 0.4, 0.4, 0.4, 0.3, 0.4, 0.1, 0.3, 0.4, 0.3 (1) As shown before, β20 = 0.3, θ20 = 0.6. Using the method in

9
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Fig. 6. The social network of 25 experts after dividing the groups.

Table 2
The weight of each expert in sub-cluster gγ .
g1
Expert e1 e2 e3 e5 e6 e7 e8 e9

Weight 0.0890 0.1189 0.1109 0.1346 0.1428 0.0967 0.1045 0.0662

g2
Expert e10 e11 e12 e13 e14 e15 e16 e17 e18 e19
Weight 0.0907 0.0813 0.1331 0.1343 0.0972 0.0667 0.0852 0.0996 0.1089 0.1031
g3
Expert e20 e21 e22 e23 e24 e25
Weight 0.2314 0.2183 0.1205 0.1278 0.1548 0.1472

Table 3
The group preference matrix.
Alternative c1 c2 c3 c4

x1 [0.6459,0.7536] [0.7202,0.8220] [0.665,0.7585] [0.5554,0.6475]

x2 [0.3089,0.3967] [0.3215,0.4420] [0.338,0.4282] [0.3954,0.6475]
x3 [0.5722,0.6660] [0.3764,0.4860] [0.480,0.5802] [0.5313,0.6322]
x4 [0.4270,0.5483] [0.4816,0.5963] [0.409,0.5161] [0.4214,0.5350]

Section 3, we can obtain T20 = {e21 , e23 , e24 }, ICI(0) (e21 ) = set I21 and the whole group at the time t = 0:
0.6596,ICI(0) (e23 ) = 0.6703, ICI(0) (e24 ) = 0.7092. The expert e24 with ⎛ ∼ ∼ ⎞
the highest consensus level in set T20 . B20 = ∅, I20 = {B20 ∩ T20 } = ∅ . ∼
3,(1)
∼
3,(0) (1 - θ21 ) ⎝A3,(0)
20 + A22
3,(0) ∼
(0) ⎠
Therefore, the expert e20 can be influenced not only by his own opinion, A21 = θ21 ⋅A21 + ⋅ + Aall
2 2
but also the opinions of the expert e24 and the whole group at the time
t = 0: Through the above analysis, we can obtain the sub-cluster consensus
∼ ∼ ( ∼ ∼ ) level, where ICI(1) (C1 ) = [0.7065,0.7070],ICI(1) (C2 ) =
3,(1) 3,(0) (1 - θ20 ) 3,(0)
[0.7042,0.7052], and ICI (C3 ) = [0.7801,0.7825], GCI(1) =
(0)
A20 = θ20 ⋅A20 + ⋅ arg max ICI (0) (A24 ) + Aall (1)
2 e24 ∈T20 ( )〈
[0.7210,0.7211]. As the consensus threshold α is 0.8 and min GCI(1) α,
(2) As shown before, β21 = 0.4, θ21 = 0.7. Using the method in the feedback mechanism is used to adjust the experts’ decision matrix.
Section 3, we can obtain T21 = {e20 , e22 , e25 }. B21 = {e20 , e22 , e23 , e24 }, Round 2:
I21 = {B21 ∩ T21 } = {e20 , e22 }. Therefore, the expert e20 can be influ­
As minICIγ(1) (Cγ ) = ICI(1) (C2 ), to be sure, experts in the sub-cluster C2
enced not only by his own opinion, but also the opinions of the expert in
need to revise their opinions according to the feedback mechanism in

10
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Table 4
The change of consensus level in each iteration.
Consensus Round 0 Round 1 Round 2 Round 3

ICI (C1)
(t) [0.7056, 0.7063] [0.7065, 0.7070] [0.7096, 0.7097] [0.8017, 0.8036]
ICI(t) (C2) [0.7032, 0.7042] [0.7042, 0.7052] [0.8080, 0.8088] [0.8103, 0.8118]
ICI(t) (C3) [0.6732, 0.6776] [0.7801, 0.7825] [0.7817, 0.7851] [0.7814, 0.7842]
GCI(t) [0.6986, 0.6988] [0.7210, 0.7211] [0.7692, 0.7696] [0.8021, 0.8027]

Section 3.2, where C2 = {e10 ,e11 ,e12 ,e13 ,e14 ,e15 ,e16 ,e17 ,e18 ,e19 }. For e10 , ∼ ∼ ( ∼ ∼ )
e11 as example: (1 - θ2 )
A1,(3)
2 = θ2 ⋅A1,(2)
2 + ⋅ A1,(2)
3 + A(2)
all
(1) As shown before, β10 = 0.1, θ10 = 0.5. Using the method in 2
Section 3, we can obtain T10 = {e11 , e12 , e13 }, ICI(1) (e11 ) = Through the above analysis, we can get the consensus level of sub-
0.6629,ICI(1) (e12 ) = 0.6893, ICI(1) (e13 ) = 0.7405. The expert e13 with clusters, where ICI(3) (C1 ) = [0.8017,0.8036],ICI(3) (C2 ) = [0.8103,
the highest consensus level in set T10 . B10 = ∅, I10 = {B10 ∩ T10 } = ∅. 0.8118], and ICI(3) (C3 ) = [0.7814,0.7842].GCI(3) = [0.8021,0.8027].
Therefore, the expert e10 can be influenced not only by his own opinion, ( )〉
As the consensus threshold α is 0.8, min GCI(3) α, consensus is reached,
but also the opinions of the expert e13 and the whole group at the time and the feedback adjustment process is over. The change of consensus
t = 1: level in each iteration is shown in Table 4.
∼ ∼
(1 - θ10 )
( ∼ ∼ ) Step 6: Obtain the most suitable alternative.
2,(2) 2,(1) 2,(1) (1)
A10 = θ10 ⋅A10 +
2
⋅ arg max ICI (1) (A13 ) + Aall
e13 ∈T10 From Eqs. (21) and (22) we can obtain the group interval number
decision matrix Aall in Table 5.
(3)
(2) As shown before, β11 = 0.4, θ11 = 0.7. Using the method in
From Eq. (35), the overall evaluation value of each alternative is
Section 3, we can obtain T11 = {e10 , e13 , e14 }. B11 =
obtained: S(x1 ) = [0.6473,0.7531], S(x2 ) = [0.3331,0.4301], S(x3 ) =
{e10 , e12 , e13 , e14 , e15 , e16 , e17 , e18 , e19 }, I11 = {B11 ∩ T11 } = {e10 , e13 , e14 }.
[0.4872,0.5919],S(x4 ) = [0.4441,0.5565].min(S(x1 )) = 0.6473>min(
Therefore, the expert e11 can be influenced not only by his own opinion,
S(x3 )) = 0.4872>min(S(x4 )) = 0.4441>min(S(x2 )) = 0.3331. Hence, the
but also the opinions of the expert in set I11 and the whole group at the
derived ranking of alternatives is x1 ≻ x3 ≻ x4 ≻ x2 . Therefore, the most
time t = 1:
suitable system is x1 .
⎛ ∼ ∼ ∼
⎞
(1 - θ11 ) ⎜A2,(1) + A2,(1) 5. Simulation analysis and comparison analyses
∼ ∼ 2,(1) ∼
13 + A14 ⎟
A2,(2)
11 = θ11 ⋅A2,(1)
11 + ⋅⎝ 10 + A(1)
all ⎠
2 3
In order to verify the validity of the proposed consensus model, we
Through the above analysis, we can obtain the sub-cluster consensus use MATLAB to simulate and compare the validity of the model.
level, where ICI(2) (C1 ) = [0.7096,0.7097],ICI(2) (C2 ) = [0.8080,0
.8088], and ICI(2) (C3 ) = [0.7817,0.7851].GCI(2) = [0.7692,0.7696]. As 5.1. Simulation analysis
( )〈
the consensus threshold α is 0.8 and min GCI(2) α, the feedback
mechanism is used to adjust experts’ decision matrix. In this section, the influence of different self-confidence values and
Round 3: bounded confidence values of experts on the consensus process reaching
As minICIγ(2) (Cγ ) = ICI(2) (C1 ), to be sure, experts in the sub-cluster C1 are explored, and the results are analyzed.
need to revise their opinions according to the feedback mechanism in Step 1: Simulation design
Section 3.2, where C1 = {e1 , e2 , e3 , e4 , e5 , e6 , e7 , e8 , e9 }. For e1 , e2 as We use the numerical example in Section 4 for analysis. Due to the
example: self-confidence value of experts belong to [0, 1]:
(1) As shown before, β1 = 0.4 ,θ1 = 0.6. Using the method in Section
3, we can obtain T1 = {e2 , e5 }. B1 = {e3 , e4 , e5 , e6 , e7 , e8 , e9 }, I1 = {B1 ∩ (1) The self-confidence value < 0.5. That means the experts are more
T1 } = {e5 }. Therefore, the expert e1 can be influenced not only by his influenced by the outside world. They believe in their own
own opinion, but also the opinions of the expert in set I1 and the whole judgment less;
group at the time t = 2: (2) The self-confidence value = 0.5. It is said that the experts are
( ∼ neutral attitude;
∼ ∼ ∼ )
1,(3)
A1 = θ1 ⋅A1 +
1,(2) (1 - θ1 ) 1,(2)
⋅ A5 + Aall
(2) (3) That the self-confidence value > 0.5 means the experts are less
2 influenced by the outside world. They are more likely to believe
(2) As shown before, β2 = 0.3 ,θ2 = 0.8. Using the method in Section in themselves. In the simulation experiment, we set the self-
3, we can obtain T2 = {e1 , e3 , e4 }. B2 = {e3 , e5 , e9 }, I2 = {B2 ∩ T2 } = confidence value ∈ {0.4, 0.5, 0.6}.
{e3 }. Therefore, the expert e2 can be influenced not only by his own
opinion, but also the opinions of the experts in set I2 and the whole group Due to the bounded confidence value of experts belonging to [0, 1]:
at the time t = 2:
(1) That the bounded confidence value = 0 means the expert is
extremely closed and does not accept the opinions of other

Table 5
The group interval number decision matrix Aall .
(3)

Alternative c1 c2 c3 c4

x1 [0.6535,0.7597] [0.7223,0.8256] [0.6452,0.7371] [0.5275,0.6507]

x2 [0.3124,0.3998] [0.3206,0.4203] [0.3163, 0.4068] [0.3997,0.5135]
x3 [0.5692,0.6725] [0.3786,0.4869] [0.4878, 0.5902] [0.5268,0.6305]
x4 [0.4302,0.5503] [0.4835,0.5897] [0.4203,0.5260] [0.4297,0.5463]

11
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Fig. 9. The average CRP iteration time at different consensus levels.

Fig. 7. Simulation experiment I. values (Fig. 8).

Step 3: Result analysis.
From Fig. 7, we can obtain the following conclusions:

(1) When the self-confidence value = 0.4, the bounded confidence

value expands from 0.05 to 0.45, and the number of iterations of
CRP decreases from 6 to 5.
(2) When the self-confidence value = 0.5, the bounded confidence
value expands from 0.05 to 0.45, and the number of iterations of
CRP remains unchanged.
(3) When the self-confidence value = 0.6, the bounded confidence
value increase from 0.05 to 0.45, and the number of iterations of
CRP decrease from 9 to 8.

In conclusion, when the self-confidence value is fixed, the greater the

bounded confidence value is, the number of iterations of CRP tends to
decrease. The fewer iterations, the more efficient the consensus is.
Therefore, the self-confidence value remains the same, and the larger the
bounded confidence value is, the more likely experts are to refer to
external opinions and reach a consensus.
From Fig. 8, we can obtain the following conclusions:

Fig. 8. Simulation experiment II. (1) When the bounded confidence value = 0.05, the self-confidence
value expands from 0.4 to 0.6, and the number of iterations of
experts who are different from his own opinion except for the CRP increases from 6 to 9.
exchange of opinions with experts who have exactly the same (2) When the bounded confidence value = 0.25, the self-confidence
opinions; value expands from 0.4 to 0.6, and the number of iterations of
(2) The bounded confidence value = 1 the expert is extremely open, CRP increases from 5 to 8.
and any other expert will become his neighbor. Meanwhile, the (3) When the bounded confidence value = 0.45, the self-confidence
opinions will be updated taking into account the opinions of any value expands from 0.4 to 0.6, and the number of iterations of
other experts. The smaller the bounded confidence value, the CRP increases from 5 to 8.
stronger the expert’s ability to understand things and the smaller
the external disturbance. In the simulation experiment, we set the In conclusion, when the bounded confidence is fixed, the larger the
bounded confidence value ∈ {0.05, 0.25, 0.45}. self-trust is, the number of iterations of CRP tends to increase. The more
iterations, the less efficient the consensus is. Therefore, the bounded
Step 2: Simulation experiment. confidence value remains the same, and the larger the self-confidence
Let the self-confidence value ∈ {0.4, 0.5, 0.6},the bounded confi­ value is, the more experts tend to believe their own judgment, and the
dence value ∈ {0.05, 0.25, 0.45}. Two simulation experiments are less they refer to external opinions, the less likely they are to reach a
designed to calculate the number of iterations when the consensus level consensus. Therefore, the proper reference to outside opinions is helpful
of the group reaches 0.9. for experts to reach consensus.
Simulation experiment I: The self-confidence value is fixed. First,
we observe the number of iterations under different bounded confidence 5.2. Comparison analyses
values (Fig. 7).
Simulation experiment II: The bounded confidence value is fixed. Next, the comparative analyses are presented to show the advantages
We observe the number of iterations under different self-confidence of the proposed method. The effectiveness and superiority of the method

12
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Table 6 alternatives by calculating geometric distance. TOPSIS can easily ex­

The ranking of the alternatives about the three methods. press human preferences, allow unlimited standards, and clarify the
Methods The ranking of the alternatives trade-offs between these standards. It can be seen from Section 4 that
when experts reach a consensus, the group decision matrix is Aall in
(3)
The proposed method x1 ≻ x3 ≻ x4 ≻ x2
x1 ≻ x3 ≻ x4 ≻ x2 Table 5. Then the overall evaluation value of each alternative is
Li et al. (2022)
obtainedS(x1 ) = [0.4340,0.4496],S(x2 ) = [0.017,0.027], S(x3 )=
TOPSIS x1 ≻ x3 ≻ x4 ≻ x2
[0.2746,0.2946],S(x4 ) = [0.2491,0.2548].min(S(x1 )) = 0.4340
>min(S(x3 )) = 0.2746 >min(S(x4 )) = 0.2491>min(S(x2 )) = 0.017.
Hence, the derived ranking of alternatives is x1 ≻ x3 ≻ x4 ≻ x2 . There­
Table 7
fore, the most suitable system is x1 . The method of Li et al. (2022) is also
The number of iterations of two methods.
used to compare the ranking of the alternatives. The ranking of the al­
The consensus threshold α The number of iterations ternatives about the three methods as follows (Table 6):
The proposed method Li et al. (2022)
What’s more, we compare the number of iterations between our
0.75 2 2 method and Li et al. (2022) when the consensus threshold is 0.75 and
0.8 3 6
0.8. The number of iterations of two methods as follows (Table 7):
At the same time, we also compare the group consensus level of the
proposed method and Li et al. (2022) with the number of iterations. The
group consensus level varies with the number of iterations as follows
(Fig. 10):
Through the above comparative analysis, the following observations
can be obtained.
It can be seen from Fig. 9 that when the consensus threshold is 0.75
and 0.8, the CRP iteration time after grouping using the Louvain method
is shorter than that without grouping using the Louvain method. This
shows that using the Louvain method to group is helpful to improve the
efficiency of large-scale group decision making. It can reduce the time it
takes experts to reach consensus.
It can be seen from Table 6 that the method proposed in this paper is
consistent with TOPSIS method and Li et al. (2022) in the ranking of the
alternatives. It shows that the proposed method is effective.
It can be seen from Table 7 that when the consensus threshold is =
0.75, both the method proposed in this paper and Li et al. (2022) require
two iterations. When the consensus threshold is 0.8, the method pro­
posed in this paper needs to be iterated 3 times, while the method of Li
et al. (2022) needs 6 times. Meanwhile, it can be seen from Fig. 10 that
the group consensus level changes with the number of iterations. When
Fig. 10. The group consensus level varies with the number of iterations. the consensus threshold is 0.8, the method proposed in this paper gets to
0.8 faster than Li et al. (2022). The method proposed in this paper takes
are compared from three aspects: CRP iteration times, the number of less time to reach the consensus threshold but Li et al. (2022) needs more
CRP iteration, and the consistency of alternative solutions. When the time to reach the consensus threshold. The results show that our method
consensus threshold is 0.75 and 0.8, we design two comparative can improve the effectiveness of large-scale group decision making.
analyses. The similarity between the method proposed in this paper and Li
First, in order to study the effect of the Louvain method on CRP, we et al. (2022) is that the Louvain algorithm is used to group experts. But
compare the CRP iteration times with and without the Louvain method. the feedback mechanisms of the two methods are different. The feed­
We ran each method 1000 times on MATLAB when the consensus back mechanism of Li et al. (2022) assumes that the experts tend to trust
threshold is 0.75 and 0.8. Their average CRP iteration times are as fol­ the opinions of the experts in the group whose opinions are less than a
lows (Fig. 9): bounded confidence value. The method in this paper assumes that the
Second, to verify the effectiveness and superiority of our proposed expert not only accepts the opinions of the experts within the group
method, the number of CRP iterations and the ranking generated by whose opinions are less than a bounded confidence value, but also tends
different methods are compared. We compare the ranking results of our to accept the opinions of the whole group. These figures show that our
method with TOPSIS method and Li et al. (2022). TOPSIS (Technique for proposed method is superior to Li et al. (2022) in decision efficiency.
Order Preference by Similarity to an Ideal Solution) method using the Therefore, it is effective to consider the opinions from these three as­
numerical example in Section 4. pects: the opinions of the experts themselves, the opinions of the people
TOPSIS was proposed by Hwang and Yoon (1981) to sort and select trusted by the experts, and the opinions of the whole group in the

Table 8
The comparison of the related literatures.
References Consensus Fuzzy preference Social network and bounded Clustering based on interval social Weights considering social
reaching process information confidence model (trust) relationships network and subgroup size

Wu et al. (2019b) × × × × ×
Xu et al. (2020) √ √ × × ×
Zheng et al. (2021) √ √ × × ×
(Zhan et al., 2022)
Li et al. (2022) √ × √ × ×
Zha et al. (2023) √ × × × ×
This paper √ √ √ √ √

13
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Table 9 reduction, we judge whether the sub-cluster reaches consensus. Other­

The list of acronyms used in this paper. wise, a tripartite feedback mechanism based on SNBC model is used to
Acronym Full Name or Explanation adjust the opinions of experts in the group with the lowest consensus
level. In order to solve the problem that the existing feedback mecha­
LSGDM Large-Scale Group Decision Making
CRP Consensus Reaching Process nism based on the bounded confidence and social network has limita­
SNBC Social Network and Bounded Confidence tions and low decision-making efficiency. Our feedback mechanism
SNA Social Network Analysis allows experts to refer not only to local opinions, but also to overall
COVID Coronavirus Disease 2019 opinions. Thus, it avoids referring only to decision opinions coming from
TOPSIS Technique for Order Preference by Similarity to an Ideal Solution
parts or only to opinions from the whole. Numerical examples, simula­
tion experiments and comparison analyses demonstrate the effective­
feedback mechanism. ness of the proposed consensus reaching model approach. By analyzing
Based on the preceding comparative analysis, we have observed that the results, we can draw the following conclusions: Using the Louvain
the method proposed in this paper outperforms other approaches in algorithm to group can effectively reduce the difficulty of LSGDM and
terms of effectiveness, efficiency, and superiority. Next, we discuss the improve the efficiency of decision-making. Considering the opinions
similarities and differences between our method and other large-scale from three aspects: the expert’s own opinion, the opinion of the person
group decision making methods(Li et al., 2022; Wu, Zhang, Liu & Cao, the expert trusts, the opinion of the group in the feedback mechanism of
2019b; Xu et al., 2020; Zha, He, Zhan & Lang, 2023; Zheng, Xu, He & CRP can effectively also improve the efficiency of decision-making.
Tian, 2021) in Table 8. Proper reference to outside opinions can help the experts reach a bet­
ter consensus.
5.3. The advantages of the proposed methods in this paper There are also some limitations to our approach: First, in the pref­
erence information structure, we use preference structure is an interval
Through the above simulation experiment and comparative analysis, number. However, experts can express their preferences not only by
the main advantages and differences of the method proposed in this interval numbers, but also by using multiple language expressions.
paper are summarized below: Second, in the social network structure, we currently study the large-
scale group decision-making consensus reaching process under static
(1) We use interval value to express social network relationships network. Third, our approach has not yet been applied to real decision
among experts and expert preferences. In large-scale group de­ problems. In view of these limitations and dilemmas, we propose
cision problems, social network relationships and preferences possible solutions and next research directions as follows: First, we will
among experts are difficult to express in precise numerical terms. study the large-scale group decision-making problem in heterogeneous
Interval value can be solved effectively the problem. language environment. Second, we will extend to study the large-scale
(2) We adopt the Louvain algorithm to divide the entire group into group decision-making consensus reaching process under a dynamic
several subgroups. The modeling calculations for each subgroup network. Third, in terms of the application context, we intend to
can be performed simultaneously, thus reducing the computation combine the method with large-scale emergency decision making
time. What’s more, the Louvain algorithm is used to cluster all methods to solve complex decision problems in uncertain environments.
experts according to a certain social network relationship. It en­
sures that each subgroup is independent but has strong cohesion. CRediT authorship contribution statement
There is a high degree of trust among experts in the same sub­
group. It can effectively solve this problem that some current Jiangjing Tan: Methodology, Software, Data curation, Writing –
clustering methods divide according to opinion similarity and original draft, Writing – review & editing. Yingming Wang: Concep­
ignore the social network relationship among experts. tualization, Visualization, Investigation. Junfeng Chu: Supervision,
(3) We establish a feedback mechanism of CRP using the SNBC model Software, Validation, Funding acquisition.
under interval number. The expert can refer not only to the expert
whose opinion is less than bounded confidence values, but also to Declaration of Competing Interest
the whole group. The numerical examples and simulation
experiment prove that our method is effective. Moreover, The authors declare that they have no known competing financial
comparative analysis shows that our method is more efficient and interests or personal relationships that could have appeared to influence
superior to the previous methods in decision efficiency. It can the work reported in this paper.
effectively manage the decision information of experts and
improve the efficiency of large-scale group decision making. Data availability

6. Conclusions The authors are unable or have chosen not to specify which data has
been used.
This paper uses the Louvain algorithm and SNBC model to design a
consensus reaching process in large scale group decision making. Acknowledgments
Our method starts by expressing the preference information and
social network relationships of the experts in terms of interval numbers. The authors thank the editor and three anonymous reviewers of this
Then uses the Louvain algorithm to divide the large group into several paper. Their insightful comments and suggestions have improved the
sub-clusters for dimension reduction. At the same time, we use the de­ quality of this paper significantly. Some remarks directly benefit from
gree centrality and closeness centrality of experts in the social network the referees’ ideas. This work was supported by the National Natural
to determine the weights of experts and groups. After dimension Science Foundation of China (NSFC) (72201066).

Appendix A

Table 9

14
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Appendix B

The interval number decision matrix of experts.

⎡ ⎤
[0.38, 0.58] [0.81, 0.91] [0.88, 0.94] [0.86, 0.96]
∼ ⎢ [0.29, 0.38] [0.86, 0.94] [0.10, 0.28] [0.27, 0.31] ⎥
A1 = ⎢ ⎥
⎣ [0.68, 0.73] [0.55, 0.64] [0.65, 0.78] [0.43, 0.50] ⎦
[0.19, 0.38] [0.13, 0.25] [0.10, 0.26] [0.31, 0.39]
⎡ ⎤
[0.81, 0.95] [0.82, 0.96] [0.83, 0.90] [0.14, 0.25]
∼ ⎢ [0.10, 0.20] [0.10, 0.25] [0.84, 0.94] [0.16, 0.20] ⎥
A2 = ⎢
⎣ [0.73, 0.88]
⎥
[0.87, 0.91] [0.68, 0.73] [0.67, 0.70] ⎦
[0.61, 0.76] [0.36, 0.56] [0.63, 0.74] [0.49, 0.66]
⎡ ⎤
[0.75, 0.81] [0.73, 0.85] [0.87, 0.97] [0.85, 0.99]
∼ ⎢ [0.19, 0.33] [0.13, 0.21] [0.19, 0.22] [0.21, 0.31] ⎥
A3 = ⎢
⎣ [0.72, 0.86]
⎥
[0.22, 0.43] [0.84, 0.92] [0.86, 0.90] ⎦
[0.21, 0.31] [0.65, 0.82] [0.71, 0.89] [0.50, 0.68]
⎡ ⎤
[0.70, 0.82] [0.18, 0.28] [0.73, 0.87] [0.82, 0.94]
∼ ⎢ [0.18, 0.26] [0.84, 0.94] [0.19, 0.21] [0.47, 0.61] ⎥
A4 = ⎢
⎣ [0.83, 0.89]
⎥
[0.20, 0.31] [0.87, 0.95] [0.72, 0.87] ⎦
[0.47, 0.54] [0.17, 0.27] [0.11, 0.19] [0.11, 0.18]
⎡ ⎤
[0.80, 0.96] [0.83, 0.98] [0.84, 0.92] [0.11, 0.21]
∼ ⎢ [0.14, 0.26] [0.13, 0.20] [0.10, 0.22] [0.59, 0.89] ⎥
A5 = ⎢
⎣ [0.31, 0.40]
⎥
[0.84, 0.98] [0.63, 0.70] [0.40, 0.51] ⎦
[0.42, 0.58] [0.60, 0.75] [0.64, 0.70] [0.43, 0.58]
⎡ ⎤
[0.80, 0.92] [0.85, 0.90] [0.13, 0.26] [0.85, 0.92]
∼ ⎢ [0.17, 0.24] [0.88, 0.92] [0.15, 0.28] [0.80, 0.93] ⎥
A6 = ⎢
⎣ [0.80, 0.91]
⎥
[0.55, 0.64] [0.86, 0.93] [0.54, 0.61] ⎦
[0.78, 0.92] [0.89, 0.94] [0.19, 0.23] [0.52, 0.62]
⎡ ⎤
[0.68, 0.83] [0.15, 0.23] [0.83, 0.92] [0.70, 0.86]
∼ ⎢ [0.81, 0.95] [0.11, 0.17] [0.11, 0.28] [0.16, 0.27] ⎥
A7 = ⎢
⎣ [0.56, 0.79]
⎥
[0.39, 0.54] [0.85, 0.92] [0.64, 0.73] ⎦
[0.47, 0.62] [0.39, 0.54] [0.26, 0.45] [0.64, 0.78]
⎡ ⎤
[0.15, 0.26] [0.86, 0.95] [0.72, 0.94] [0.86, 0.97]
∼ ⎢ [0.81, 0.91] [0.35, 0.45] [0.13, 0.20] [0.31, 0.44] ⎥
A8 = ⎢
⎣ [0.88, 0.90]
⎥
[0.23, 0.36] [0.75, 0.84] [0.37, 0.57] ⎦
[0.37, 0.51] [0.42, 0.58] [0.20, 0.39] [0.13, 0.20]
⎡ ⎤
[0.83, 0.96] [0.88, 0.99] [0.85, 0.92] [0.12, 0.28]
∼ ⎢ [0.17, 0.24] [0.82, 0.99] [0.24, 0.39] [0.15, 0.26] ⎥
A9 = ⎢
⎣ [0.83, 0.84]
⎥
[0.67, 0.72] [0.86, 0.96] [0.62, 0.75] ⎦
[0.26, 0.47] [0.16, 0.21] [0.10, 0.24] [0.71, 0.82]
⎡ ⎤
[0.52, 0.66] [0.84, 0.94] [0.87, 0.97] [0.82, 0.90]
∼ ⎢ [0.11, 0.18] [0.18, 0.25] [0.17, 0.20] [0.12, 0.22] ⎥
A10 = ⎢
⎣ [0.19, 0.24]
⎥
[0.12, 0.20] [0.16, 0.29] [0.25, 0.30] ⎦
[0.12, 0.29] [0.29, 0.47] [0.24, 0.36] [0.25, 0.40]
⎡ ⎤
[0.15, 0.28] [0.73, 0.83] [0.14, 0.27] [0.70, 0.84]
∼ ⎢ [0.21, 0.29] [0.10, 0.20] [0.22, 0.27] [0.11, 0.25] ⎥
A11 = ⎢
⎣ [0.66, 0.77]
⎥
[0.14, 0.24] [0.50, 0.65] [0.31, 0.46] ⎦
[0.68, 0.77] [0.68, 0.75] [0.56, 0.59] [0.57, 0.65]
⎡ ⎤
[0.88, 0.90] [0.84, 0.96] [0.10, 0.23] [0.67, 0.79]
∼ ⎢ [0.18, 0.20] [0.16, 0.20] [0.82, 0.95] [0.87, 0.97] ⎥
A12 = ⎢
⎣ [0.23, 0.30]
⎥
[0.43, 0.59] [0.18, 0.27] [0.49, 0.54] ⎦
[0.23, 0.34] [0.73, 0.89] [0.54, 0.62] [0.30, 0.48]
⎡ ⎤
[0.80, 0.90] [0.71, 0.82] [0.88, 0.93] [0.14, 0.26]
∼ ⎢ [0.88, 0.93] [0.19, 0.25] [0.17, 0.25] [0.87, 0.98] ⎥
A13 = ⎢
⎣ [0.85, 0.95]
⎥
[0.12, 0.29] [0.12, 0.22] [0.44, 0.56] ⎦
[0.80, 0.93] [0.79, 0.85] [0.60, 0.74] [0.49, 0.60]

15
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

⎡ ⎤
[0.55, 0.65] [0.83, 0.90] [0.89, 0.99] [0.11, 0.27]
∼ ⎢ [0.20, 0.25] [0.14, 0.20] [0.50, 0.61] [0.13, 0.23] ⎥
⎢
A14 = ⎣ ⎥
[0.11, 0.20] [0.49, 0.52] [0.41, 0.57] [0.31, 0.43] ⎦
[0.18, 0.24] [0.58, 0.69] [0.22, 0.31] [0.58, 0.71]
⎡ ⎤
[0.11, 0.29] [0.75, 0.91] [0.81, 0.91] [0.80, 0.97]
∼ ⎢ [0.13, 0.20] [0.14, 0.22] [0.19, 0.24] [0.22, 0.30] ⎥
⎢
A15 = ⎣ ⎥
[0.30, 0.48] [0.18, 0.28] [0.51, 0.59] [0.26, 0.33] ⎦
[0.85, 0.96] [0.69, 0.77] [0.26, 0.33] [0.39, 0.49]
⎡ ⎤
[0.44, 0.53] [0.55, 0.70] [0.88, 0.95] [0.70, 0.85]
∼ ⎢ [0.15, 0.25] [0.15, 0.23] [0.83, 0.96] [0.18, 0.28] ⎥
⎢
A16 = ⎣ ⎥
[0.66, 0.72] [0.11, 0.25] [0.33, 0.49] [0.26, 0.35] ⎦
[0.79, 0.85] [0.57, 0.67] [0.48, 0.52] [0.33, 0.42]
⎡ ⎤
[0.88, 0.98] [0.61, 0.72] [0.29, 0.32] [0.70, 0.88]
∼ ⎢ [0.17, 0.28] [0.64, 0.78] [0.14, 0.21] [0.12, 0.24] ⎥
=
A17 ⎣⎢ ⎥
[0.12, 0.22] [0.15, 0.29] [0.21, 0.33] [0.38, 0.54] ⎦
[0.71, 0.80] [0.69, 0.70] [0.60, 0.78] [0.28, 0.32]
⎡ ⎤
[0.40, 0.50] [0.87, 0.98] [0.29, 0.36] [0.84, 0.91]
∼ ⎢ [0.11, 0.26] [0.65, 0.86] [0.17, 0.26] [0.10, 0.20] ⎥
=
A18 ⎣⎢ ⎥
[0.82, 0.94] [0.53, 0.64] [0.34, 0.45] [0.67, 0.74] ⎦
[0.55, 0.67] [0.15, 0.24] [0.70, 0.86] [0.64, 0.79]
⎡ ⎤
[0.81, 0.99] [0.84, 0.92] [0.82, 0.93] [0.21, 0.30]
∼ ⎢ [0.86, 0.90] [0.15, 0.20] [0.17, 0.20] [0.11, 0.25] ⎥
=
A19 ⎣⎢ ⎥
[0.17, 0.28] [0.14, 0.24] [0.13, 0.20] [0.50, 0.63] ⎦
[0.18, 0.25] [0.63, 0.77] [0.31, 0.40] [0.54, 0.66]
⎡ ⎤
[0.13, 0.23] [0.83, 0.95] [0.85, 0.93] [0.83, 0.98]
∼ ⎢ [0.14, 0.22] [0.11, 0.27] [0.82, 0.97] [0.16, 0.25] ⎥
=
A20 ⎣⎢ ⎥
[0.43, 0.52] [0.82, 0.96] [0.15, 0.23] [0.81, 0.96] ⎦
[0.14, 0.29] [0.11, 0.23] [0.16, 0.20] [0.37, 0.48]
⎡ ⎤
[0.85, 0.95] [0.95, 0.99] [0.85, 0.94] [0.11, 0.23]
∼ ⎢ [0.20, 0.35] [0.15, 0.25] [0.85, 0.94] [0.79, 0.86] ⎥
=
A21 ⎣⎢ ⎥
[0.65, 0.70] [0.35, 0.41] [0.39, 0.46] [0.68, 0.78] ⎦
[0.15, 0.39] [0.30, 0.44] [0.12, 0.27] [0.58, 0.62]
⎡ ⎤
[0.89, 0.92] [0.81, 0.95] [0.80, 0.92] [0.10, 0.24]
∼ ⎢ [0.17, 0.21] [0.82, 0.94] [0.24, 0.34] [0.80, 0.91] ⎥
=
A22 ⎣⎢ ⎥
[0.64, 0.76] [0.17, 0.29] [0.51, 0.62] [0.87, 0.92] ⎦
[0.17, 0.29] [0.38, 0.52] [0.32, 0.47] [0.15, 0.22]
⎡ ⎤
[0.89, 0.94] [0.12, 0.20] [0.80, 0.90] [0.81, 0.94]
∼ ⎢ [0.11, 0.25] [0.10, 0.22] [0.10, 0.22] [0.80, 0.90] ⎥
=
A23 ⎣⎢ ⎥
[0.87, 0.95] [0.16, 0.24] [0.32, 0.47] [0.59, 0.75] ⎦
[0.22, 0.40] [0.12, 0.28] [0.17, 0.20] [0.14, 0.20]
⎡ ⎤
[0.89, 0.99] [0.86, 0.90] [0.17, 0.20] [0.84, 0.90]
∼ ⎢ [0.17, 0.24] [0.11, 0.21] [0.19, 0.24] [0.77, 0.81] ⎥
=
A24 ⎣⎢ ⎥
[0.77, 0.83] [0.18, 0.21] [0.81, 0.91] [0.84, 0.97] ⎦
[0.48, 0.51] [0.23, 0.35] [0.78, 0.90] [0.63, 0.78]
⎡ ⎤
[0.81, 0.99] [0.25, 0.35] [0.88, 0.94] [0.10, 0.20]
∼ ⎢ [0.80, 0.92] [0.14, 0.25] [0.16, 0.20] [0.10, 0.24] ⎥
=
A25 ⎣⎢ ⎥
[0.88, 0.90] [0.57, 0.62] [0.52, 0.67] [0.37, 0.42] ⎦
[0.18, 0.25] [0.65, 0.72] [0.79, 0.85] [0.20, 0.34]

References Zhang, C. H., Su, W. H., Zeng, S. Z., Balezentis, T., & Herrera-Viedma, E. (2021). A two-
stage subgroup decision-making method for processing large-scale information.
Expert Systems with Applications, 171, Article 114586.
Pelissari, R., Oliveira, M. C., Ben Amor, S., Kandakoglu, A., & Helleno, A. L. (2020). Smaa
Zhong, X. Y., Xu, X. H., & Pan, B. (2022). A non-threshold consensus model based on the
methods and their applications: A literature review and future research directions.
minimum cost and maximum consensus-increasing for multi-attribute large group
Ann. Oper. Res., 293(2), 433–493.
decision-making. Information Fusion, 77, 90–106.
Qin, J. D., Li, M. X., & Liang, Y. Y. (2022). Minimum cost consensus model for crp-driven
Chen, X. H., Zhang, W. W., Xu, X. H., & Cao, W. Z. (2022a). A public and large-scale
preference optimization analysis in large-scale group decision making using louvain
expert information fusion method and its application: Mining public opinion via
algorithm. Information Fusion, 80, 121–136.
sentiment analysis and measuring public dynamic reliability. Information Fusion, 78,
71–85.

16
J. Tan et al. Expert Systems With Applications 237 (2024) 121560

Ding, R. X., Palomares, I., Wang, X. Q., Yang, G. R., Liu, B. S., Dong, Y. C., et al. (2020). Zha, Q. B., Liang, H. M., Kou, G., Dong, Y. C., & Yu, S. (2019). A feedback mechanism
Large-scale decision-making: Characterization, taxonomy, challenges and future with bounded confidence- based optimization approach for consensus reaching in
directions from an artificial intelligence and applications perspective. Information multiple attribute large-scale group decision-making. IEEE Transactions on
Fusion, 59, 84–102. Computational Social Systems, 6(5), 994–1006.
Dong, Y. C., Zha, Q. B., Zhang, H. J., Kou, G., Fujita, H., Chiclana, F., et al. (2018). Zhang, Z., Gao, Y., & Li, Z. (2020). Consensus reaching for social network group decision
Consensus reaching in social network group decision making: Research paradigms making by considering leadership and bounded confidence. Knowledge-Based
and challenges. Knowledge-Based Systems, 162, 3–13. Systems, 204, Article 106240.
Du, Z. J., Luo, H. Y., Lin, X. D., & Yu, S. M. (2020). A trust-similarity analysis-based Liu, P., Li, Y., & Wang, P. (2023). Opinion dynamics and minimum adjustment-driven
clustering method for large-scale group decision-making under a social network. consensus model for multi-criteria large-scale group decision making under a novel
Information Fusion, 63, 13–29. social trust propagation mechanism. IEEE Transactions on Fuzzy Systems, 31(1),
Zhong, X. Y., Xu, X. H., & Yin, X. P. (2021). A multi-stage hybrid consensus reaching 307–321.
model for multi-attribute large group decision-making: Integrating cardinal Li, Y. H., Kou, G., Li, G. X., & Peng, Y. (2022). Consensus reaching process in large-scale
consensus and ordinal consensus. Computers & Industrial Engineering, 158, Article group decision making based on bounded confidence and social network. European
107443. Journal of Operational Research, 303(2), 790–802.
Herrera-Viedma, E., Palomares, I., Li, C. C., Cabrerizo, F. J., Dong, Y. C., Chiclana, F., Urena, R., Chiclana, F., Melancon, G., & Herrera-Viedma, E. (2019). A social network
et al. (2021). Revisiting fuzzy and linguistic decision making: Scenarios and based approach for consensus achievement in multiperson decision making.
challenges for making wiser decisions in a better way. IEEE Transactions on Systems Information Fusion, 47, 72–87.
Man Cybernetics-Systems, 51(1), 191–208. Liang, W., Labella, Á., Wang, Y.-M., & Rodríguez, R. M. (2023). Consensus reaching
Cheng, D., Yuan, Y. X., Wu, Y., Hao, T. T., & Cheng, F. X. (2022). Maximum satisfaction process under interval-valued hesitant fuzzy environment. Computers & Industrial
consensus with budget constraints considering individual tolerance and compromise Engineering, 176, Article 108971.
limit behaviors. European Journal of Operational Research, 297(1), 221–238. Wu, Y. Z., Zhang, Z., Kou, G., Zhang, H. J., Chao, X. R., Li, C. C., et al. (2021). Distributed
Dong, Y. C., Zha, Q. B., Zhang, H. J., & Herrera, F. (2021). Consensus reaching and linguistic representations in decision making: Taxonomy, key elements and
strategic manipulation in group decision making with trust relationships. IEEE applications, and challenges in data science and explainable artificial intelligence.
Transactions on Systems Man Cybernetics-Systems, 51(10), 6304–6318. Information Fusion, 65, 165–178.
Deng, X. Y., Xue, S. Y., & Jiang, W. (2023). A novel quantum model of mass function for Zhou, M., Chen, Y. W., Liu, X. B., Cheng, B. Y., & Yang, J. B. (2020). Weight assignment
uncertain information fusion. Information Fusion, 89, 619–631. method for multiple attribute decision making with dissimilarity and conflict of
Wu, T., & Liu, X. (2023). Large-scale group decision-making with uncertain and behavioral belief distributions. Computers & Industrial Engineering, 147, Article 106648.
considerations: Methods and applications. Springer. Zhou, M., Guan, Z. X., Chen, Y. W., Zhou, Z. P., Wu, J., & Herrera-Viedma, E. (2022).
Zhang, H. J., Dong, Y. C., & Herrera-Viedma, E. (2018). Consensus building for the Consistency and consensus reaching process for group decision making based on
heterogeneous large-scale gdm with the individual concerns and satisfactions. IEEE complete interval distributed preference relations under social network analysis.
Transactions on Fuzzy Systems, 26(2), 884–898. Information Fusion, 88, 126–145.
Wu, T., Liu, X. W., Qin, J. D., & Herrera, F. (2019a). Consensus evolution networks: A Ma, X. J., Gong, Z. W., Wei, G., & Herrera-Viedma, E. (2022). A new consensus model
consensus reaching tool for managing consensus thresholds in group decision based on trust interactive weights for intuitionistic group decision making in social
making. Information Fusion, 52, 375–388. networks. IEEE Rransactions on Cybernetics, 52(12), 13106–13119.
Chu, J. F., Wang, Y. M., Liu, X. W., & Liu, Y. C. (2020). Social network community Herrera-Viedma, E., Alonso, S., Chiclana, F., & Herrera, F. (2007). A consensus model for
analysis based large-scale group decision making approach with incomplete fuzzy group decision making with incomplete fuzzy preference relations. IEEE Transactions
preference relations. Information Fusion, 60, 98–120. on Fuzzy Systems, 15(5), 863–877.
Wu, T., Liu, X., Qin, J., & Herrera, F. (2022). A new clustering algorithm with preference Kumar, K., & Chen, S.-M. (2022). Group decision making based on improved linguistic
adjustment cost to reduce the cooperation complexity in large-scale group decision interval-valued atanassov intuitionistic fuzzy weighted averaging aggregation
making. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 52(8), operator of linguistic interval-valued atanassov intuitionistic fuzzy numbers.
5271–5283. Information Sciences, 607, 884–900.
Liu, X., Xu, Y. J., & Herrera, F. (2019a). Consensus model for large-scale group decision Chen, X., Wang, X., Zhang, H. B., Xu, Y. H., Chen, Y., & Wu, X. T. (2022b). Interval topsis
making based on fuzzy preference relation with self-confidence: Detecting and with a novel interval number comprehensive weight for threat evaluation on
managing overconfidence behaviors. Information Fusion, 52, 245–256. uncertain information. Journal of Intelligent & Fuzzy Systems, 42(4), 4241–4257.
Opsahl, T., Agneessens, F., & Skvoretz, J. (2010). Node centrality in weighted networks: Kacprzyk, J. (1986). Group decision-making with a fuzzy linguistic majority. Fuzzy Sets
Generalizing degree and shortest paths. Social Networks, 32(3), 245–251. and Systems, 18(2), 105–118.
Tian, Z. P., Nie, R. X., & Wang, J. Q. (2019). Social network analysis-based consensus- Kay, E. (1977). Graph theory with applications. Journal of the Operational Research
supporting framework for large-scale group decision-making with incomplete Society, 28(1), 237–238.
interval type-2 fuzzy information. Information Sciences, 502, 446–471. Wasserman, S., & Faust, K. (1994). Structural analysis in the social sciences. Social network
Liu, B. S., Zhou, Q., Ding, R. X., Palomares, I., & Herrera, F. (2019b). Large-scale group analysis: Methods and applications. Cambridge University Press.
decision making model based on social network analysis: Trust relationship-based Sabidussi, G. (1966). The centrality of a graph. Psychometrika, 31(4), 581–603.
conflict detection and elimination. European Journal of Operational Research, 275(2), Chengzhong, L. (1989). Fuzzy set theory.Beijing. Beijing Normal University Press.
737–754. You, T. H., & Fan, Z. P. (2002). Topsis method for multiple attribute decision making
Li, Y. H., Kou, G., Li, G. X., & Wang, H. M. (2021). Multi-attribute group decision making with intervals. Journal of Northeastern University, 23(9), 840–843.
with opinion dynamics based on social trust network. Information Fusion, 75, Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of
102–115. communities in large networks. Journal of Statistical Mechanics: Theory and
Tan, X., Zhu, J. J., Cabrerizo, F. J., & Herrera-Viedma, E. (2021). A cyclic dynamic trust- Experiment, 2008(10), P10008.
based consensus model for large-scale group decision making with probabilistic Waltman, L., & van Eck, N. J. (2013). A smart local moving algorithm for large-scale
linguistic information. Applied Soft Computing, 100, Article 106937. modularity-based community detection. European Physical Journal B, 86(11), 471.
Xu, X. H., Zhang, Q. H., & Chen, X. H. (2020). Consensus-based non-cooperative Hwang, C. L., & Yoon, K. P. (1981). Multiple attribute decision making: Methods and
behaviors management in large-group emergency decision-making considering applications: A state-of-the-art survey. SpringerVerlag.
experts’ trust relations and preference risks. Knowledge-Based Systems, 190, Article Wu, T., Zhang, K., Liu, X., & Cao, C. (2019b). A two-stage social trust network partition
105108. model for large-scale group decision-making problems. Knowledge-Based Systems,
Hassani, H., Razavi-Far, R., Saif, M., Chiclana, F., Krejcar, O., & Herrera-Viedma, E. 163, 632–643.
(2022). Classical dynamic consensus and opinion dynamics models: A survey of Zha, Q., He, X., Zhan, M., & Lang, N. (2023). Managing consensus in balanced networks
recent trends and methodologies. Information Fusion, 88, 22–40. based on opinion and trust/distrust evolutions. Information Sciences, 643, Article
Zhang, B. W., Dong, Y. C., Zhang, H. J., & Pedrycz, W. (2020). Consensus mechanism 119223.
with maximum-return modifications and minimum-cost feedback: A perspective of Zheng, Y., Xu, Z., He, Y., & Tian, Y. (2021). A hesitant fuzzy linguistic bi-objective
game theory. European Journal of Operational Research, 287(2), 546–559. clustering method for large-scale group decision-making. Expert Systems with
Hegselmann, R., & Krause, U. (2002). Opinion dynamics and bounded confidence Applications, 168, Article 114355.
models, analysis and simulation. Journal of Artificial Societies and Social.
SIMULATION, 5.

17