Professional Documents
Culture Documents
Abstract—The current societal demands and technological Index Terms—Compatibility index, conflict, fuzzy c-means
developments have resulted in the participation of a large number clustering (FCM), multicriteria large-scale group decision making
of experts in making decisions as a group. Conflicts are immi- (LSGDM), subgroup-to-subgroup.
nent in groups and conflict management is complex and necessary
especially in a large group. However, there are few studies that
quantitatively research the conflict detection and resolution in the
large-group context, especially in the multicriteria large-group I. I NTRODUCTION
decision making (GDM) context. This article proposes a dynamic ROUP decision making (GDM) can be defined as a situa-
adaptive subgroup-to-subgroup conflict model to solve multicrite-
ria large-scale GDM problems. A compatibility index is proposed
G tion in which two or more experts make their evaluations
toward a set of possible alternatives and then select the optimal
based on two kinds of conflicts among experts: 1) cognitive
conflict and 2) interest conflict. Then, the fuzzy c-means cluster- one [1]. In GDM, experts have their own attitudes, motiva-
ing algorithm is used to classify experts into several subgroups. tions, and perceptions. Hence, they tend to have different
A subgroup-to-subgroup conflict detection method and a weight- opinions on available alternatives. Conflicts among experts are
determination approach are developed based on the clustering caused because of such discrepancies in opinions. Conflict
results. Afterward, a conflict resolution model, which can dynam-
ically generate feedback suggestion, is introduced. Finally, an is widely defined as a struggle or clash that occurs when
illustrative example is provided to demonstrate the effectiveness a perceived incompatibility or threat exists in the thoughts,
and applicability of the proposed model. behaviors, feelings, or attitudes of two or more parties [2].
How to achieve high group effectiveness and obtain high-
quality decision results through effective conflict management
Manuscript received July 28, 2019; revised December 21, 2019; accepted
February 15, 2020. This work was supported in part by the National Natural has been a topic of common concern for organizational man-
Science Foundation of China under Grant 71771156, and in part by Spanish agers and researchers [3]. However, most existing studies used
Ministry of Science and Universities Funds under Grant TIN2016-75850-R. empirical research [4] or meta-analysis [5] to test the relation-
This article was recommended by Associate Editor Q. Shen. (Corresponding
authors: Huchang Liao; Enrique Herrera-Viedma.) ship between conflict and team performance. Few studies built
Ming Tang is with the Business School, Sichuan University, Chengdu models to measure the degree of group conflict quantitatively
610064, China (e-mail: tangming0716@163.com). and introduce interactive feedback methods to solve conflicts
Huchang Liao is with the Business School, Sichuan University, Chengdu
610064, China, and also with the Faculty of Computing and Information among experts.
Technology, King Abdulaziz University, Jeddah 21589, Saudi Arabia (e-mail: With the rapid expansion of technological paradigms, such
liaohuchang@163.com). as e-democracy [6] and social networks [7], there is a trend
Enrique Herrera-Viedma is with the Andalusian Research Institute in
Data Science and Computational Intelligence, University of Granada, that a larger number of experts are involved in the decision-
18071 Granada, Spain, and also with the Department of Electrical and making process. Therefore, the large-scale GDM (LSGDM)
Computer Engineering, Faculty of Engineering, King Abdulaziz University, becomes a much-talked-about topic. Many challenges stem-
Jeddah 21589, Saudi Arabia (e-mail: viedma@decsai.ugr.es).
C. L. Philip Chen is with the Department of Computer and Information ming from general GDM to LSGDM will arise, such as
Science, Faculty of Science and Technology, University of Macau, Macau, the supervision cost [8], social relationship [9], noncoop-
China, also with the Navigation College, Dalian Maritime University, erative behavior [10], minority opinion [11], and conflict
Dalian 116026, China, and also with the Unmanned System Research
Institute, Northwestern Polytechnical University, Xi’an 710072, China (e-mail: among experts. However, as an important research issue
philip.chen@ieee.org). in the field of GDM, conflict analysis is relatively less
Witold Pedrycz is with the Department of Electrical and Computer researched by scholars in LSGDM problems. When the
Engineering, University of Alberta, Edmonton, AB T6G 2R3, Canada, and
also with the Department of Electrical and Computer Engineering, Faculty of number of experts in a group is large, there will be
Engineering, King Abdulaziz University, Jeddah 21589, Saudi Arabia (e-mail: many more conflicts than conventional GDM problems.
wpedrycz@ualberta.ca). Consequently, it is difficult to measure and manage con-
Color versions of one or more of the figures in this article are available
online at http://ieeexplore.ieee.org. flicts in LSGDM problems. If conflicts cannot be coordinated
Digital Object Identifier 10.1109/TCYB.2020.2974924 properly, then the time cost will increase significantly, and
2168-2267
c 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
the decision-making efficiency will decrease, especially in the 2) A new integrated weight-determination method, which
LSGDM environment. combines opinion transition probability of a subgroup
To detect and manage conflicts, it is necessary to deter- to another and the size of subgroups, is proposed. The
mine the types of conflicts. The 2-D model is the most opinion transition probability is calculated based on the
widely used principle to identify the types of conflicts [4]. Markov chain.
Different types under this principle, such as the task rela- 3) A dynamic adaptive subgroup-to-subgroup conflict
tionship, cognitive relationship, and cognitive interest have model is developed. The feedback mechanism can gen-
been predominantly researched by scholars [4]. The dimen- erate suggestions dynamically according to the conflict
sional of cognitive interest was proposed by Doucet et al. degree of each subgroup against other subgroups in each
[12]. Cognitive conflicts appear because different experts have iterative round.
various knowledge levels or experiences, and therefore make This article is organized as follows. In Section II, the
different evaluations on the same thing. Interest conflicts description of the multicriteria LSGDM problem is given.
appear because different experts represent different interest Section III introduces the conflict detection process in detail.
groups, and therefore make different evaluations from their In Section IV, we discuss how to resolve conflicts for
own interests. For a multicriteria LSGDM problem, the cog- multicriteria LSGDM problems. An illustrative example is
nitive interest is mainly reflected by different evaluation values provided to verify the applicability of the proposed model
of alternatives, while the interest conflict is mainly reflected by in Section V. Some discussions regarding the proposed
the inconsistency of the importance for each criterion [13]. Let model are given in Section VI. Conclusions are drawn
us use a real-world example to illustrate this issue. The loca- in Section VII.
tion of a large shopping mall is a complex decision-making
problem, which often involves large-scale experts and officials
from multiple fields and departments, such as the municipal
II. D ESCRIPTION OF THE M ULTICRITERIA
department, land and resources, environmental protection, traf-
LSGDM P ROBLEM
fic, and gardening and greening. Multiple criteria should be
considered, such as the cost, economical and population char- A multicriteria LSGDM problem is defined as a situation in
acteristics, environmental consideration, transportation consid- which at least 20 experts are involved in the decision-making
eration, and investor’s competency. When evaluating possible problem to select an optimal alternative considering multiple
alternatives, experts from different departments provide their criteria. The main elements of a multicriteria LSGDM problem
evaluation information according to their knowledge. However, include the following:
when weighting criteria, different departments may have their 1) a set of alternatives, X = {x1 , x2 , . . . , xm }, m ≥ 2, which
emphases from their interests. For instance, the traffic depart- are selected as possible solutions to the problem;
ment focuses on transportation consideration since it does not 2) a set of experts, E = {e1 , e2 , . . . , eQ }, Q ≥ 20, which
want the mall to significantly increase the traffic pressure are invited to evaluate alternatives and give their assess-
around it. ments; in an LSGDM problem, there are two widely
To reduce the scale of LSGDM problems, the clustering used principles for the number of experts [13]: one is
process is the most popular [8]. It can classify all experts 20 [8]; and the other is 11 [11]; this article uses the
into several subgroups, and then the size and complexity former principle;
of the problem can be reduced. In such a case, the sub- 3) a set of criteria, C = {c1 , c2 , . . . , cn }, n ≥ 2, and the
of criteria ω = (ω1 , ω2 , . . . , ωn ) , with
weight vector T
group becomes a decision unit in the LSGDM. We can
measure the discrepancy between a subgroup and the collective ωj ≥ 0 and nj=1 ωi = 1.
group. Most previous LSGDM studies used the centricity- Each expert expresses his/her opinions on alternatives over
q q
oriented method of adopting the aggregated group opinion criteria by a decision matrix Vq = (vij )m×n , where vij denotes
as a reference point to find incompatible subgroups [13]. the assessment of alternative xi on criterion cj . In this arti-
However, this method has some limitations. First, the pre- cle, we assume that all experts use the scale 0–1 to provide
condition of the centricity-oriented method is that the aggre- evaluation values.
gation function is defined. But there are many aggregation Formally, the solving process for the multicriteria LSGDM
functions, and different aggregation functions may lead to problem usually requires four steps [11].
different results. In this sense, the robustness of the centricity- Step 1: Normalize original decision matrices. To elim-
oriented method is weak. Second, the required aggrega- inate the influence of dimension, all assessments should
tion step increases the complexity and cost of LSGDM be normalized into values in the interval [0, 1]. For
problems. a multicriteria LSGDM problem, generally, the criteria can
Bearing all aforementioned points in mind, this article be classified into two categories: 1) benefit criteria and
develops a conflict detection and resolution model for mul- 2) cost criteria. In this article, we transform the indi-
q
ticriteria LSGDM problems. Theoretically, this article has the vidual decision matrix Vq = (vij )m×n into the nor-
q q
following contributions. malized decision matrix Dq = (dij )m×n , where dij =
q q q q
1) Two kinds of conflicts are defined for multicrite- [(vij − mini {vij })/(maxi {vij } − mini {vij })] for benefit crite-
q q q q q
ria LSGDM problems. The compatibility index is rion cj , and dij = [(maxi {vij } − vij )/(maxi {vij } − mini {vij })],
developed based on these two kinds of conflicts. for cost criterion cj .
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TANG et al.: DYNAMIC ADAPTIVE SUBGROUP-TO-SUBGROUP COMPATIBILITY-BASED CONFLICT DETECTION AND RESOLUTION MODEL 3
TABLE I
S YMBOL D ESCRIPTIONS a clustering algorithm is adopted to classify experts into
several subgroups. Based on the clustering results, a subgroup-
to-subgroup conflict detection method and a weight-
determination method are developed.
1 q
n
ICqs = ωj − ωjs (3)
n
j=1
q
where ωj and ωjs are the weights of the jth criterion given by
experts eq and es , respectively.
Both CCqs and ICqs belong to [0, 1]. The larger CCqs or
ICqs is, the greater the cognitive or interest conflict degree
Step 2: Classify experts into several subgroups. In this between eq and es .
step, we use the fuzzy c-means clustering (FCM) algo- We use the 2-D principle to define these two kinds of con-
rithm to classify all experts {e1 , e2 , . . . , eQ } into K subgroups flicts. Note that there may be also other conflicts that cannot
G1 , G2 , . . . , GK . The weight vector of criteria and the decision be included by these two kinds of conflicts such as the incon-
matrix corresponding to the kth subgroup are denoted as ωGk sistency in screening indicators. These conflicts can influence
G
and DGk = (dij k )m×n , respectively. the decision-making process. In this article, only cognitive and
Step 3: Aggregate all subgroups’ decision matrices into interest conflict are considered.
a global decision matrix. Note that this aggregation pro- Inspired by Saaty [14] and Chen et al. [15], we then define
cess is only applied after the conflict threshold is reached. the compatibility index between two experts.
q q q
Suppose that (λ1 , λ2 , . . . , λK )T is the weight vector of all sub- Definition 1: Let ωq = (ω1 , ω2 , . . . , ωn )T and ωs =
groups. The weighted averaging (WA) operator can be used to (ω1 , ω2 , . . . , ωn ) be the weight vectors of criteria given by
s s s T
q
G the experts eq and es , respectively. Let Dq = (dij )m×n and
aggregate DGk = (dij k )m×n into a collective decision matrix
Ds = (dij )m×n be the decision matrices provided by the experts
s
DG = (dijG )m×n , such that
eq and es , respectively. The compatibility index of eq and es
K is defined as
G ⎛ ⎞
dijG = WA dijG1 , dijG2 , . . . , dijGK = λk dij k (1) n n
1 ζ m
q q
k=1 c eq , es = ⎝ dij − dijs + (1 − ζ ) ωj − ωjs ⎠
n m
where λk is the weight of Gk and satisfies λk ≥ 0,
i=1 j=1 j=1
K
k=1 λk = 1. (4)
Step 4: Select the best alternative. This step obtains the final
where ζ is a parameter to control the attention on the interest
evaluation value of each alternative using the WA operator.
conflict and cognitive conflict. If ζ > 0.5, then (4) pays
For better understanding, we summarize mathematical sym-
more attention to the impact of cognitive conflict; if ζ > 0.5,
bols used in this article in Table I.
then (4) pays more attention to the impact of interest conflict;
ζ = 0.5 denotes that (4) pays equal attention to two kinds of
III. C ONFLICT D ETECTION IN M ULTICRITERIA LSGDM conflicts. In this article, we set ζ as 0.5.
In this section, we introduce the cognitive conflict and Through this process, the discrepancies between experts
interest conflict in multicriteria LSGDM problems. Then, with regard to the weights of criteria and evaluation values in
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TANG et al.: DYNAMIC ADAPTIVE SUBGROUP-TO-SUBGROUP COMPATIBILITY-BASED CONFLICT DETECTION AND RESOLUTION MODEL 5
G G
(dij k )m×n and DGl = (dij l )m×n be the decision matrix of sub- Gk may have a similar opinion with subgroup Gl and have
group Gk and Gl , respectively. The compatibility index of Gk a diverse opinion from subgroup Gp . The differences in the
and Gl can be defined as closeness degrees between subgroups reveal the possibilities of
⎛ opinion transitions from one subgroup to another. Generally,
1 ⎝ ζ Gk
m n
G people do the same thing with other people who have similar
c(Gk , Gl ) = dij − dij l
n m opinions with them [26]. Thus, it is easier to let Gk modify
i=1 j=1
⎞ its opinion more similar to Gl ’s opinion than to Gp ’s because
n
Gk G of the higher closeness degree of their opinions.
+ (1 − ζ ) ωj − ωj l ⎠. (10) Moreover, opinion transition probabilities reveal the agree-
j=1 ment level between one subgroup to another. This pro-
According to (10), we have c(Gk , Gl ) ∈ [0, 1] and cess of opinion transition can be regarded as a finite
c(Gk , Gl ) = 0 if and only if DGk = DGl and ωGk = ωGl . state space Markov chain [27]. Inspired by Ross [28] and
Next, we can define the subgroup conflict index. Dong et al. [25], we develop a method to obtain the weight
Definition 3: Let ωGk , ωGl , DGk , and DGl be as before. vector based on opinion transition probabilities. We define the
The subgroup conflict index (SCI) between Gk and Gl is opinion transition probability from one subgroup to another
defined as based on the subgroup conflict index SCI
1/SCIkl
SCIkl = c(Gk , Gl ). (11) pkl = K 1
(12)
l=1 SCIkl
SCI holds: 1) SCIkl ∈ [0, 1]; 2) SCIkl = SCIlk (symme- K
try); and 3) SCIkk = 0 (reflexivity). Based on (11), we can where l pkl = 1, pkl > 0. A Markov chain matrix
obtain the subgroup conflict matrix O = (SCIkl )K×K , where P = (pkl )K×K can be derived by computing opinion transi-
the elements are conflict degrees between pairwise subgroups. tion probabilities between any pair of subgroups. The weight
A reasonable threshold ψ(0 < ψ < 1) should be set to distribution of subgroups can be obtained by the limiting distri-
confirm whether there is a conflict between two subgroups. bution which is the only stationary distribution of the Markov
For Gk and Gl , if SCIkl > ψ, then they are in conflict. chain. Ross [28] provided a method to calculate the limiting
Otherwise, if SCIkl ≤ ψ, then the divergences of assess- distribution π
ments between Gk and Gl are not enough to create a conflict.
πTP = π (13)
Obviously, it is unrealistic and unnecessary to let SCIkl = 0
K
in real-world decision-making problems. A certain degree of where π = (π1 , π2 , . . . , πK )T , k πk = 1, πk ≥ 0. Let
conflict can be beneficial for group performance because it γk = πk , then the weight vector γ 1 = (γ11 , γ21 , . . . , γK1 )T
1
promotes the exploration and exchanging of different ideas, can be derived.
stimulates in-depth analysis of alternative action plans, and The size is also considered regarding the importance of
effectively inhibits groupthink [23], [24]. The value of ψ can a subgroup. Let γk2 = #Gk /Q, then the weight vector
be determined according to specific situations. A vital and γ 2 = (γ12 , γ22 , . . . , γK2 )T can be derived. The final weight of a
major problem should set a low conflict threshold such as 0.1. subgroup is
At the other extreme, such as urgent problems, the conflict
γ1 × γ2
threshold should be high because of the time limitation. The γk = K k 1 k 2 . (14)
conflict threshold should also take into account the group size. k=1 γk × γk
An LSGDM problem should have a higher conflict threshold
The final weight vector of subgroups is γ = (γ1 , γ2 , . . . , γK )T .
than a small-scale GDM problem since more conflicts exist in
Next, we analyze the time complexity of the conflict detec-
LSGDM problems.
tion process. First, we calculate the compatibility degree
between each pair of experts based on (2)–(4). The time com-
D. Weight Determination for Subgroups plexity is O((mn+n+1)×Q(Q − 1)/2), where Q is the number
The conflict detection and resolution model proposed in of experts involved in the problem, m is the number of alter-
this article does not have the aggregation process. It relies natives, and n is the number of criteria. The time complexity
on pairwise comparisons between subgroups, instead of the of the FCM clustering algorithm is O(2Q2 K 2 i) [29], where 2
general centricity-oriented weight determination method that is the dimension of data and i is the number of iterations. In
computes the similarity degree between subgroups and the the clustering process, we added a selection method for initial
aggregated group. Therefore, it seems of interest to develop clustering centroids. We select the point that has the smallest
a weight determination method that adapts to the context of average distance with other points as the first clustering cen-
the subgroup-to-subgroup conflict model. In this regard, this troid. Then, K points can be selected using this method. The
article proposes an integrated method combining the Markov time complexity is no more than O(KQ). Similar to (4), the
chain and the size of subgroups. time cost of (10) is O((mn + n + 1) × K(K − 1)/2). The time
Based on (10), we can obtain the degree of conflict cost of (12) and (14) is O(K(K −1)/2+K +1) and O(2K +1),
for each pair of subgroups. Then, their opinion transition respectively. There are also other calculation processes, which
probabilities [25] can be derived. For a large group of experts, can be implemented offline. Therefore, the time complexity of
there is usually a diversity of assessments. However, subgroup these processes is ignored.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
IV. A DAPTIVE C ONFLICT R ESOLUTION M ODEL more changes should be made in its opinions. K Furthermore,
The conflict coordination and resolution is an important αkt = αlt , if and only if K y=1,y=k,l SCIky = y=1,y=k,l SCIly .
research topic in LSGDM. If there are pairs of subgroups The subgroup conflict index can only reveal the undi-
that do not reach the conflict threshold, then, the feedback rected relations between pairs of subgroups for the reason that
mechanism should be applied to reduce the conflict degree. SCIkl = SCIlk (symmetry). However, for a pair of subgroups
{Gk , Gl }, there may be different adoption directions between
these two subgroups because they can take different attitudes
A. Adaptive Mechanism to Reduce the Conflict Degree of toward the adjusted suggestions. Therefore, the ordered pair
Subgroups of subgroups are defined to clarify the direction between two
Suppose that Gk and Gl are selected in the t-th iteration to subgroups.
make modifications according to feedback suggestions. The Definition 4: Let G = {G1 , G2 , . . . , GK } be a set of K sub-
t+1
new weight vector ωG k
and the decision matrix Dt+1
Gk can be groups. An ordered pair of subgroups is defined as (Gk , Gl ).
modified by the following strategies: (Gk , Gl ) denotes that subgroup Gk may revise its opinions
α t 1−α t according to subgroup Gl .
ωG t+1
= ω t
Gk
k
ωGt k
(15) Based on Definition 4, we can find that (Gk , Gl ) = (Gl , Gk ).
k l
α t 1−α t The details of the conflict resolution model are described as
t+1 l l
ωG l
= ω t
Gl ω t
Gk (16) follows.
α t 1−α t Note: There are two strategies to revise experts’
t+1 k k
dij(G k)
= d t
ij(Gk ) d t
ij(Gl ) (17) opinions [20]: one is the automatic strategy [30], which mod-
α t 1−α t ifies assessment values automatically without experts’ inter-
t+1 l l
dij(G l)
= d t
ij(Gl ) d t
ij(Gk ) (18) actions. The other is the feedback strategy which provides
suggestions by a moderator for experts and requires them to
where αkt is a parameter that determines what amounts of make modifications. The automatic strategy is time consum-
t+1 t+1 ing because it does not have the interactive process between
new opinions (ωG k
and dij(G k)
) of Gk are from original opin-
t t
ions (ωGk and dij(Gk ) ). As we can see from (15)–(18), each experts and the moderator. However, the automatic strategy
new opinion of a subgroup contains two parts: its own orig- makes the decision-making result only a calculated one. One
inal opinion and the original opinion of another subgroup main problem of the feedback strategy is that experts may
who is the most compatible one with it in the entire group. reject suggestions. Furthermore, the moderator may bring sub-
After this process, the opinions of two subgroups will come jectivity. The proposed approach is a comprehensive method
closer in the next iteration. The new weight vector is ωG t+1
= that combines two strategies. The experts only provide their
t+1 t+1 t+1 n t+1
k
initial judgments in the initial iterative round, and the mecha-
(ω1(Gk ) , ω2(Gk ) , . . . , ωn(Gk ) ). Note that j=1 ωj(Gk ) = 1.
nism generates suggestions automatically in the next iterative
Therefore, we need to normalize it.The normalized weight
n rounds. The experts only need to determine to accept the
vector of criteria is ωt+1 t+1
j(Gk ) = ωj(Gk ) /
t+1
j=1 ωj(Gk ) .
suggestions or not.
It is important to determine the value of αkt because it
Theorem 1: In the t-th round of CRP, 0 ≤ t ≤ t∗ − 1, if
reflects the portion of opinion preservation in each iterative
SCItkl ≥ ψ ∀ ∈ k, l ∈ E, we have
round. αkt can be set in advance or determined dynamically
according to the specific situation of each iteration. In this max SCIt+1 ≤ max SCItkl . (25)
kl
article, we determine αkt dynamically. k,l∈E k,l∈E
In the conflict resolution process (CRP), it is reasonable The proof of Theorem 1 is given in the Appendix.
that the subgroups with higher incompatibility with other sub-
groups need to make more modifications than those subgroups
B. Algorithm of the Adaptive Conflict Resolution Model for
that with lower conflict degrees. Therefore, for each subgroup,
Multicriteria LSGDM Problems
the conflict resolution model provides a dynamic αkt changing
according to the conflict degree with unselected subgroups, Next, we present the algorithm of the proposed adaptive
such that conflict resolution model, which is summarized as follows.
K The flow of Algorithm 3 is displayed in Fig. 1.
y=1,y=k,l SCIky For Algorithm 3, we have analyzed the main computational
αk = 1 −
t
K (19)
K cost of steps 1–3 and 5 in Section III. For the CRP, the
2 y=1,y=k,l SCIky + y=1,y=k,l SCIly
K computational cost in the worst case is O(2(n + mn)T).
y=1,y=k,l SCIly
αl = 1 −
t
K . (20)
2 K
SCI + SCI V. I LLUSTRATIVE E XAMPLE : E NERGY C ENTER
y=1,y=k,l ky y=1,y=k,l ly
S ELECTION OF THE D ISTRIBUTED E NERGY S YSTEM
We can see that αkt , αlt ∈ (0.5, 1). This can avoid exces- A. Example Description
sive changes of original opinions andenhance the cred- The distributed energy system (DES) is relative to the
K
ibility of decision-making results. If
K y=1,y=k,l SCIky ≥ traditional centralized energy supply system. The traditional
y=1,y=k,l SCI ly , then αk
t ≤ αl
t . It denotes that if a subgroup centralized energy supply system adopts large capacity equip-
has a higher conflict degree with unselected subgroups, then ment and centralized production, and then transmits various
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TANG et al.: DYNAMIC ADAPTIVE SUBGROUP-TO-SUBGROUP COMPATIBILITY-BASED CONFLICT DETECTION AND RESOLUTION MODEL 7
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TABLE II
O RIGINAL W EIGHTS OF C RITERIA AND D ECISION M ATRICES
B. Solution Process
To obtain the optimal alternative, the following steps are
given based on the proposed adaptive conflict resolution
model. In this article, we set ψ = 0.1, T = 10, and ξ = 0.001.
Step 1: Calculate the compatibility degree of experts. To
save space, the compatibility matrix is presented in the
Appendix.
Step 2: Use the FCM clustering algorithm to classify
experts. By (9), we can obtain the relationship between
the number of clusters and value of L, which is shown in
Fig. 2. Therefore, the optimal number of clusters is 4. The
clustering result is provided in Table III.
The subgroup conflict matrix is
⎛ ⎞
− 0.1528 0.1671 0.2181
⎜ 0.1528 − 0.1153 0.1403 ⎟
O0 = SCI0kl =⎜ ⎝
⎟.
K×K 0.1671 0.1153 − 0.1759 ⎠
0.2181 0.1403 0.1759 −
Since the threshold of the conflict degree is not reached, a subgroup accept). Then, we have
the CRP is adopted to decrease the conflict degree. We have ω1G1 = (0.2782, 0.4326, 0.2892)T
SCI014 = maxG0 {SCI0kl }. Based on (19) and (20), we have
α10 = 0.7485 and α40 = 0.7515. Then, the automatic feedback ω1G4 = (0.3284, 0.4826 , 0.1890)T
⎛ ⎞
mechanism is used to generate modification suggestions for 0.5341 0.4511 0.5239
G1 and G4 based on (15)–(18). Suppose that both subgroups dG1
1
= ⎝ 0.5563 0.6000 0.2320 ⎠
accept the suggestions (no less than two thirds of experts in 0.5926 0.6037 0.1786
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TANG et al.: DYNAMIC ADAPTIVE SUBGROUP-TO-SUBGROUP COMPATIBILITY-BASED CONFLICT DETECTION AND RESOLUTION MODEL 9
TABLE III
S TRUCTURE OF S UBGROUPS The threshold of the conflict level is still not reached with
t = 2 < T and G2 = ∅. From O2 , we have SCI213 =
maxG2 {SCI2kl }. Based on (19) and (20), we have α12 = 0.7437
and α32 = 0.7563. Then, the automatic feedback mechanism is
implemented to provide modification suggestions for G2 and
G3 based on (15)–(18). Suppose that both subgroups accept
the suggestions. Then, we have
ω3G1 = (0.3046, 0.4084, 0.2808)T
ω3G3 = (0.3659, 0.3674, 0.2667)T
⎛ ⎞
0.5201 0.4570 0.4945
dG3
1
= ⎝ 0.5057 0.5374 0.2808 ⎠
0.5614 0.5702 0.2290
⎛ ⎞
0.4937 0.4686 0.4418
dG3
3
= ⎝ 0.4200 0.4333 0.4073 ⎠
0.5051 0.5100 0.3718
O3 = SCI3kl
⎛ K×K
⎞
− 0.1042 0.0564 0.0950
⎜ 0.1042 − 0.1001 0.1075 ⎟
=⎜⎝ 0.0564 0.1001
⎟.
− 0.0938 ⎠
0.0950 0.1075 0.0938 −
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TABLE IV
C OMPARISON A MONG T HREE M ETHODS such as “low,” “medium,” or “high,” and generated feedback
suggestions. Our proposed model can provide feedback sug-
gestions according to specific values of conflict degrees. This
mechanism is achieved based on the dynamic parameter α.
VI. D ISCUSSION
In this section, we provide some discussions regarding the
proposed conflict model. The main features of this model can
be summarized as follows.
C. Comparative Study 1) Unlike most existing LSGDM models, the proposed con-
To clearly differentiate our model from others, in this sec- flict resolution model is constructed based on subgroup-
tion, we provide a comparative study from three perspectives: to-subgroup comparisons. Our model does not need
1) weight determination method; 2) conflict measure; and the reference point. The centricity-oriented method may
3) adaptive CRP. lead to biases because different aggregation functions
1) Weight Determination Method: Many existing studies may cause different results. Furthermore, in the multi-
used the majority principle [11]. The weight of a subgroup criteria LSGDM environment, it is an arduous task to
is determined by its size. However, this principle has limi- aggregate all experts’ opinions in each iterative round.
tations. Two subgroups with the same number of members Therefore, our proposed model provides a new frame-
but different inner characteristics will have the same weight. work to solve multicriteria LSGDM problems. The
To overcome this limitation, Rodríguez et al. [31] provided biases and costs in the aggregation process can be
an integrated method that combines the size and cohesion of reduced.
a subgroup. Our method combines the majority principle and 2) In each iterative round, the experts involved in the
the degree of conflict. A subgroup with a low degree of conflict feedback mechanism and the amount of adjusted opin-
with other subgroups will be given a large weight. Therefore, ions are different. These two points correspond to the
it can accelerate the decision-making process to some extent. word “dynamic” in the title. Furthermore, if a subgroup
2) Conflict Measure: There are two widely used methods has a higher conflict degree with other subgroups, then
to measure the degree of discrepancy of a group. The first more changes will be made in its opinions. This point
(M1) is based on the distance to the collective opinion, that corresponds to the word “adaptive” in the title.
is, the centricity-oriented method. This method first aggregates 3) The feedback mechanism considers the situation in
all subgroups’ opinions into a group collective opinion. Then, which experts may reject the modification suggestions.
each subgroup’s discrepancy degree can be obtained by cal- The feedback suggestions are generated automatically
culating the distance between a subgroup and the collective without the participation of experts. However, the gen-
group. The limitations of this method have been analyzed in erated suggestions will return to experts to keep them
Section I. The second method is based on the distance of informed. The subgroup can choose to accept or reject
subgroups. This method (M2) calculates the similarity degree the suggestions. In other words, this mechanism can
and obtains the similarity matrices of each pair of subgroups. guarantee the rights of rejection of experts.
Afterward, the consensus matrix is generated by aggregating In addition to these advantages, we can also find limitations,
all similarity matrices. The discrepancy degree in the level which require future research efforts.
of alternatives and subgroups can be obtained according to 1) The accuracy of the FCM clustering algorithm remains
the consensus matrix. Our proposed method does not need an issue. A good clustering result usually means a high
to generate the consensus matrix and analyze the discrepancy degree of intracluster similarity and a low degree of
for pairs of subgroups directly. Therefore, the computational intercluster similarity. Many external and internal cri-
process is simplified. A brief summarization of these three teria have been developed to evaluate the accuracy of
methods is given in Table IV. clustering results [34]. It seems of interest to research
3) Adaptive CRP: There are also other adaptive models. this issue and use these criteria to investigate the struc-
For instance, Mata et al. [32] distinguished the level of con- tures of subgroups.
sensus into three situations: very low, low, and medium, and 2) In some extreme situations, experts could not find one or
then proposed different strategies to deal with these situations. more suitable criteria. In such a case, they may prefer
The number of experts that need to make modifications and not to provide assessment information on the weights
the number of changes on opinions were adaptively different of criteria. Thus, managing incomplete information
in three situations. Rodríguez et al., [31] applied this idea to and implementing psychology concepts or persuasion
the LSGDM context. Pérez et al. [33] distinguished experts principles to convince experts to provide assessment
into three categories with respect to their weights: 1) high- information are promising topics. Furthermore, it is
importance experts; 2) medium-importance experts; and 3) also possible that a subgroup has a significant conflict
low-importance experts, and then generated different feed- with other subgroups. This involves the topic of minor-
back strategies adaptively according to experts’ weights. As ity opinions, as mentioned in Section I. Some studies
we can see, these models only distinguished several situations, have focused on minority opinions [10], [11]. How to
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
TANG et al.: DYNAMIC ADAPTIVE SUBGROUP-TO-SUBGROUP COMPATIBILITY-BASED CONFLICT DETECTION AND RESOLUTION MODEL 11
TABLE V
C OMPATIBILITY M ATRIX
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
Thus [13] M. Tang, and H. C. Liao, “From conventional group decision making
to large-scale group decision making: What are the challenges and how
to meet them in big data era? A state-of-the-art survey,” Omega, to be
max SCIt+1 kl = max max SCI t+1
kl , max SCI t+1
yk , published, doi: 10.1016/j.omega.2019.102141.
k,l∈E k,l∈E,k,l=y,z k∈E,k=y
[14] T. L. Saaty, “A ratio scale metric and the compatibility of ratio scales:
max SCIt+1 zk , SCIt+1
yz
The possibility of arrow’s impossibility theorem,” Appl. Math. Lett.,
k∈E,k=z vol. 7, no. 6, pp. 45–49, Nov. 1994.
[15] H. Y. Chen, L. G. Zhou, and B. Han, “On compatibility of uncertain
< max max SCItkl , SCItyz additive linguistic preference relations and its application in the group
k,l∈E,k,l=y,z decision making,” Knowl. Based Syst., vol. 24, no. 6, pp. 816–823,
= SCItyz = max SCItkl . Aug. 2011.
k,l∈E [16] J. C. Bezdek, R. Ehrlich, and F. Wull, “FCM: The fuzzy c-means
clustering algorithm,” Comput. Geosci., vol. 10, nos. 2–3, pp. 191–203,
2) One subgroup accepts while the other rejects. We 1984.
assume that Gy accepts. We have [17] H.-X. Pei, Z.-R. Zheng, C. Wang, C.-N. Li, and Y.-H. Shao, “D-FCM:
Density based fuzzy c-means clustering algorithm with application
in medical image segmentation,” Procedia Comput. Sci., vol. 122,
max SCIt+1 kl = max max SCIt+1
kl , max SCIt+1 yk
pp. 407–414, 2017.
k,l∈E k,l∈E,k,l=y k∈E [18] Z. Cebeci and F. Yildiz, “Comparison of k-means and fuzzy c-means
algorithms on different cluster structures,” J. Agr. Inf., vol. 6, no. 3,
< max max SCItkl , SCItyz pp. 13–23, 2015.
k,l∈E,k,l=y [19] N. R. Pal and J. C. Bezdek, “On cluster validity for the fuzzy c-means
= SCIyz = max SCItkl .
t model,” IEEE Trans. Fuzzy Syst., vol. 3, no. 3, pp. 370–379, Aug. 1995.
k,l∈E [20] M. Tang, X. Y. Zhou, and H. C. Liao, “Ordinal consensus measure
with objective threshold for heterogeneous large-scale group decision
t+1
3) Both Gy and Gz reject. Then, we have ωG = making,” Knowl. Based Syst., vol. 180, pp. 62–74, Sep. 2019.
k [21] I. Katsavounidis, C. C. J. Kuo, and Z. Zhang, “A new initialization
t+1 t+1
ωGk and dij(Gk ) = dij(Gk ) . Thus, maxk,l∈E {SCIkl } ≤
t t
technique for generalized Lloyd iteration,” IEEE Signal Process. Lett.,
maxk,l∈E {SCItkl }. This completes the proof. vol. 1, no. 10, pp. 144–146, Oct. 1994.
[22] X. L. Xie and G. Beni, “A validity measure for fuzzy clustering,”
IEEE Trans. Pattern Anal. Mach. Intell., vol. 13, no. 8, pp. 841–847,
Aug. 1991.
[23] M. Loughry and A. C. Amason, “Why won’t task conflict cooperate?
R EFERENCES Deciphering stubborn results,” Int. J. Conflict Manag., vol. 25, no. 4,
pp. 333–358, 2014.
[1] J. D. Qin, X. W. Liu, and W. Pedrycz, “An extended TODIM multi- [24] L. Pelled, K. M. Eisenhardt, and K. R. Xin, “Exploring the black box: An
criteria group decision making method for green supplier selection in analysis of work group diversity, conflict, and performance,” Administ.
interval type-2 fuzzy environment,” Eur. J. Oper. Res., vol. 258, no. 2, Sci. Quart., vol. 44, no. 1, pp. 1–28, Mar. 1999.
pp. 626–638, Apr. 2017. [25] Q. X. Dong and O. Cooper, “A peer-to-peer dynamic adaptive consensus
[2] M. Deutsch, The Resolution of Conflict: Constructive and Destructive reaching model for the group AHP decision making,” Eur. J. Oper. Res.,
Processes, New Haven, CT, USA: Yale Univ. Press, 1973. vol. 250, no. 2, pp. 521–530, Apr. 2016.
[3] A. C. Amason, K. R. Thompson, W. A. Hochwarter, and A. W. Harrison, [26] I. J. Pérez, F. J. Cabrerizo, A. Alonso, Y. C. Dong, F. Chiclana, and
“Conflict: An important dimension in successful management teams,” E. Herrera-Viedma, “On dynamic consensus processes in group decision
Organ. Dyn., vol. 24, no. 2, pp. 20–35, Aug. 1995. making problems,” Inf. Sci., vol. 459, pp. 20–35, Aug. 2018.
[4] T. A. O’Neill and M. J. W. McLarnon, “Optimizing team conflict [27] G. Tauchen, “Finite state markov-chain approximations to univariate and
dynamics for high performance teamwork,” Human Resour. Manag. Rev., vector autoregressions,” Econ. Lett., vol. 20, no. 2, pp. 177–181, 1986.
vol. 28, no. 4, pp. 378–394, Dec. 2018. [28] S. M. Ross, Introduction to Probability Models. Boston, MA, USA:
[5] C. K. W. De Dreu and L. R. Weingart, “Task versus relationship conflict, Academic, 2014.
team performance, and team member satisfaction: A meta-analysis,” [29] A. Rui and J. M. C. Sousa, “Comparison of fuzzy clustering algorithms
J. Appl. Psychol., vol. 88, no. 4, pp. 741–749, Aug. 2003. for classification,” in Proc. Int. Symp. Evolving Fuzzy Syst., Sep. 2006,
[6] R. Efremov, D. R. Insua, and A. Lotov, “A framework for participatory pp. 112–117.
decision support using Pareto frontier visualization, goal identifica- [30] Z. M. Zhang and W. Pedrycz, “A consistency and consensus-based goal
tion and arbitration,” Eur. J. Oper. Res., vol. 199, no. 2, pp. 459–467, programming method for group decision-making with interval-valued
Dec. 2009. intuitionistic multiplicative preference relations,” IEEE Trans. Cybern.,
[7] B. S. Liu, Q. Zhou, R.-X. Ding, I. Palomares, and F. Herrera, “Large- vol. 49, no. 10, pp. 3640–3654, Oct. 2019.
scale group decision making model based on social network analysis: [31] R. M. Rodríguez, Á. Labella, G. D. Tré, and L. Martínez, “A large scale-
Trust relationship-based conflict detection and elimination,” Eur. J. Oper. consensus reaching process managing group hesitation,” Knowl. Based
Res., vol. 275, no. 2, pp. 737–754, Jun. 2019. Syst., vol. 159, pp. 86–97, Nov. 2018.
[8] M. Tang, H. C. Liao, J. P. Xu, D. Streimikiene, and X. S. Zheng, [32] F. Mata, L. Martínez, and E. Herrera-Viedma, “An adaptive consensus
“Adaptive consensus reaching process with hybrid strategies for large- support model for group decision-making problems in a multigranu-
scale group decision making,” Eur. J. Oper. Res., vol. 282, no. 3, lar fuzzy linguistic context,” IEEE Trans. Fuzzy Syst., vol. 17, no. 2,
pp. 957–971, May 2020. pp. 279–290, Apr. 2009.
[9] R. X. Ren, M. Tang, and H. C. Liao, “Managing minority opinions in [33] I. J. Pérez, F. J. Cabrerizo, A. Alonso, and E. Herrera-Viedma,
micro-grid planning by a social network analysis-based large scale group “A new consensus model for group decision making problems with non-
decision making method with hesitant fuzzy linguistic information,” homogeneous experts,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 44,
Knowl. Based Syst., vol. 189, Feb. 2020, Art. no. 105060. no. 4, pp. 494–498, Apr. 2014.
[10] I. Palomares, L. Martínez, and F. Herrera, “A consensus model to [34] M. Halkidi, Y. Batistakis, and M. Vazirgiannis, “On clustering valida-
detect and manage noncooperative behaviors in large-scale group deci- tion techniques,” J. Intell. Inf. Syst., vol. 17, nos. 2–3, pp. 107–145,
sion making,” IEEE Trans. Fuzzy Syst., vol. 22, no. 3, pp. 516–530, Dec. 2001.
Jun. 2014. [35] H. C. Liao, Z. S. Xu, E. Herrera-Viedma, and F. Herrera, “Hesitant fuzzy
[11] X. H. Xu, Z. J. Du, and X. H. Chen, “Consensus model for linguistic term set and its application in decision making: A state-of-the
multi-criteria large-group emergency decision making considering non- art survey,” Int. J. Fuzzy Syst., vol. 20, no. 7, pp. 2084–2110, 2018.
cooperative behaviors and minority opinions,” Decis. Support Syst., [36] H. C. Liao, X. M. Mi, and Z. S. Xu, “A survey of decision-
vol. 79, pp. 150–160, Nov. 2015. making methods with probabilistic linguistic information: Bibliometrics,
[12] O. Doucet, J. Poitras and D. Chénevert, “The impacts of leadership on preliminaries, methodologies, applications and future directions,”
workplace conflicts,” Int. J. Conflict Manag., vol. 20, no. 4, pp. 340–354, Fuzzy Optimi. Decis. Making, vol. 19, no. 1, pp. 81–134, 2020,
2009. doi: 10.1007/s10700-019-09309-5.
Authorized licensed use limited to: University of Electronic Science and Tech of China. Downloaded on March 15,2020 at 11:57:35 UTC from IEEE Xplore. Restrictions apply.