Knowledge-Based Systems 203 (2020) 106158

Contents lists available at ScienceDirect

Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys

Recognition of opinion leaders coalitions in online social network

using game theory
∗
Lokesh Jain a , Rahul Katarya a , , Shelly Sachdeva b
a
Department of Computer Science & Engineering Delhi Technological University, New Delhi, India
b
Department of Computer Science & Engineering National Institute of Technology, New Delhi, India

article info a b s t r a c t

Article history: Nowadays, human decision-making power and attitude revolutionize over time and change their
Received 27 February 2020 viewpoint towards substantial convinced values because of the influence of social media. Social
Received in revised form 8 May 2020 network plays a crucial role in human life with the progression of technology. In the online social
Accepted 16 June 2020
network, the responsibility of opinion leaders is also vital in information diffusion and promotion
Available online 20 June 2020
purposes. In this manuscript, we proposed a different Game theory-based Opinion Leader Detection
Keywords: (GOLD) algorithm to identify the group of users having the maximum synergy declared as the coalition
Game theory of opinion leaders. For implementing purpose, we used the game theory approach in which all
Opinion leader the users behave like a player, and a trustor–trustee tree formed comprised coalitions based on
Online Social Network trust and conditional probability. We proposed an inventive and distinctive solution to measures
Shapley value the individual payoff using the distance-based centrality parameter. We also computed the Shapley
value for each user to identify the maximum marginal contribution and determined the maximum
synergy of each coalition. The main advantage of the proposed approach is to strengthen the power
of the coalition and produced the synergetic outcome. The proposed approach delivered around 90%
accuracy and decreased 37% execution time. Therefore, it provides superior results regarding the
accuracy, precision, time complexity, rate of convergence, and computational time with other SNA
(Social Network Analysis) measures.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction
In real life, a human cannot survive without any communi-
cation and interaction. Despite lots of technological inventions
and innovations, the whole existence depends on that interaction.
As the era changed, the massive growth in various technical and
social interaction mediums happened; still, the communication
among the users do not discontinue. The online social network
played a vital role in effective communication among users [1].
Users share their views and ideas on various topics related to pol-
itics, economics, current affairs, marketing, technologies, sports,
entertainment, and various similar issues. The interactions among
the users are useful to find the opinion of the group. Sometimes
it is also benefited the industry to promote its products. Some-
times, the interaction among the users affects the behavior and
decision-making process of another process [2–4].
Nowadays, the social network provides an opportunity for
users to cooperate with other users to share information. Due
∗ Corresponding author.
E-mail addresses: nsitlokesh.jain@gmail.com (L. Jain),
rahulkatarya@dtu.ac.in (R. Katarya), shellysachdeva@nitdelhi.ac.in (S. Sachdeva).
to this cooperation, there are some set of people who can influ-
ence or change the behavior and assessment of another person
known as opinion leader [5–7]. The assignment to identifying
such influential and aberrant people is known as opinion leader
detection in the social network. Sometimes, opinion leaders also
referred to as ‘trendsetter’, ‘influencer’, ‘opinion maker’, ‘thought
leader’, and ‘thought maker’. The opinion leader may be clas-
sified in various categories: (i) local and global based on the
scope of influence, i.e., global opinion leader can influence a
dense and large amount of community. A local opinion leader
only influences disjointed communities, (ii) monomorphic and
polymorphic based on operational area. Monomorphic opinion
leaders work in a specific area. At the same time, the polymor-
phic opinion leader acts in a multidisciplinary field. Generally,
polymorphic opinion leader has significant influence and impact
on target community due to their versatile range of knowledge,
(iii) positive and destructive opinion leader based on the positive
or negative power of influence and leadership. It observed that
all the opinion leaders do not work on the positive side of the
impact or information dissemination. Sometimes, opinion leaders
also intend towards downgrading the reputation of a product
and broadcast some negative information in the community be-
cause of marketing strategies. Such types of opinion leaders are

https://doi.org/10.1016/j.knosys.2020.106158
0950-7051/© 2020 Elsevier B.V. All rights reserved.
2 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
Fig. 1. A synergetic coalition of nodes (v2 , v7 , v11 ).
self-centered, narrow-minded, and parochial, (iv) short-term and
long-term opinion leaders based on the time interval in which
they affected the followers. We consider the impact of the opinion
leader’s influence over the period. The domination of opinion
leaders affects the decision of followers because the follower’s
preference also depends on the opinion leader’s trust and belief
dynamics [8]. The province of opinion leaders includes politi-
cal science, health, finance, business and marketing, information
diffusion, education, economics, and many more [9,10].
We hypothesized that game theory played a significant role
in finding out the opinion leader in the social network. Accord-
ing to the game theory approach, the action performed by one
user affects the outcomes of another user that they intended to
achieve. A game theory approach is logical learning of the strate-
gic interaction of coherent users [11]. In the real-life, the users do
not know whether their actions affect the other user’s outcomes,
but they must understand that the other user knows their efforts.
The primary assumption in the game theory is that the user must
be rational and attempted to maximize their payoff. The social
network also adopted the game theory approach, in the same
way, to find the opinion leader in promoting and capitalizing
on the growth of products in the real-world commercial mar-
ket [12,13]. Game theory also approaches beneficial to identify
the different types of power centrality and relationships between
the users in the social network [14,15]. Various issues related
to network privacy and security also resolved by game theory
approaches [16].
For instance, consider a network with a total of 16 nodes, as
shown in Fig. 1. The top nodes with higher degree centrality are
v5 , v7 , v8 , v11 , and v14 . Here, the node v2 with higher between-
ness centrality, and node v11 with higher closeness centrality;
but it is not the right approach to analyze the node’s importance
based on a single attribute because only one or two nodes are
not responsible for the information diffusion and set of nodes
disseminate the information in the network.
We observed that the synergetic coalition of nodes (v2 , v7 , v11 )
could spread or stop the information dissemination depending
upon circumstances. If any of the false information flooded by
any other node, the coalition of these nodes can easily shatter
the propagation of information.
An opinion leader is a commanding and powerful human being
or collection of people who can transform the other’s decision-
making process by their technical and intellectual abilities [17].
Nowadays, most of the companies hire the opinion leaders to
promote their products in the market. Opinion leaders are con-
structive for the diffusion of new products [18,19]. As the era
of social media revolutionized, most of the marketing strategies
have shifted to social media sites. Therefore, an enormous virtual
communication established between the consumers and the com-
panies with the expansion of social networking sites, micro-blogs,
blogs, etc. [9,20].
In this article, we handled the dilemma of opinion leader
detection in the online social network with the game theory ap-
proach. Although there is no standard and conventional method
to identify the opinion leader due to the dynamic nature of the
social network, yet we made an effort to identify the opinion
leaders in the online social network. The main benefit of the
game theory approach is to provide a logical resolution in a
competitive situation. It is also helpful to analyze and evalu-
ate the feedback of the users about any product in the net-
work. Game theory also provides more certain and concise in-
formation about human decisions using different adaptable and
optimal strategies [21]. In this research, we have proposed a
unique Game theory-based Opinion Leader Detection (GOLD) al-
gorithm to uniquely discover the cluster of opinion leaders in the
social network based on synergy and marginal payoff. The main
contribution of this investigation is to deploy the game theory
approach in the field of social network rationally.
Moreover, to calculate the individual payoff, we have proposed
the four alternative solutions. All the solution has their signif-
icance for the different types of social network and measured
over different parameters such as distance and degree of influ-
ence. We have also considered and weighed the three different
types of trust; direct, indirect, and recommended to calculate the
individual trust score. Next, the trust score used to define the
probability of the eagerness of the user to establish a connection
with each other. The proposed approach is very much practical
for the large scale social network because as the total number
of users increases, the probability of making the connections
also increases. After all, the users have more options to establish
coalitions with other users. Also, this approach does not leverage
data mining techniques, which are not too efficient to handle
such type of dilemma due to lots of statistical computation. The
proposed algorithm experimented on two real-world data sets:
Wiki-Vote social network and Bitcoin OTC trust weighted signed
network and got around 90% accuracy and 94% precision for
both the datasets. Similarly the execution time is also reduced
to about 37%, and the total number of opinion leaders required
for diffusion purposes also decreased.
The skeleton of the paper is as follows; in the first section of
the article, we dealt with the literature survey and related work
to opinion leader detection. In the next section, we elaborated
on the background information of game theory, social network
analysis, and Shapley value used for organizing our research. In
the third section, we illustrated the proposed methodology and
the two algorithms; the first algorithm measured the degree of
trust while the second algorithm identifying the group of opinion
leaders based on maximum synergy. We also provided the four
alternative solutions to calculate the payoff of the individual
users. In the next segment, we elaborated on the introduction
of datasets, and their and experimental results include the list of
opinion leaders. We also compared the derived outcomes with
the other SNA measures to validate the proposed approach and
found that the proposed algorithm provided better results in
respect of the accuracy, precision, recall, F1-score, and execution
time [22]. In the last section, we explained the conclusion, lim-
itations, and future perspective of the paper. We also discussed
some potential applications and opportunities to define the appli-
cability of opinion leaders with some additional features. Hence,
the principal contribution of this article is as follows:
• A unique algorithm (GOLD) is proposed to identify the coali-
tions.
• A new algorithm addressed to measure direct, indirect, and
recommended trust.
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
• Four alternative solutions anticipated to calculate the indi-
vidual payoff.
• Shapley value used to measure the average marginal payoff
of the user.
• Produced improved outcomes and reduced execution time
over SNA measures.
2. Related work
With the advancement of technologies, lots of approaches
have proposed to identify the opinion leader in the online social
network. The technological movement and revolution play a cru-
cial position to identify the opinion leaders in the social network.
An opinion leader can be detected qualitatively and quantitatively
based on the domain of interest need. The conditions for choosing
the opinion leaders may vary in each method. However, we
analyzed and investigated different approaches based on various
parameters such as centrality, power, trust, responsiveness, social
status, knowledge, retweets, topic sensitivity, posted text content,
meta-heuristic algorithms, and many more to classify the opinion
leaders. The summary of the intact literature papers, along with
their benefits and limitations, are shown in Table 1.
If we observed the above table gravely, we found that different
researches utilized the various parameters and centrality mea-
sures to choose the list of opinion leaders. Generally, the selection
of opinion leaders broadly classified into two categories based
on the network’s inherent characteristics and the user’s behavior
in the network, respectively. The first category depends on the
network’s formation, dimension, density, and degree of a node.
On the other hand, the second one builds upon the user’s retweet,
response time, trust, test mining, and user’s behavior. Besides the
mentioned literature, some researcher has different outlooks and
ways of network formation to recognize opinion leaders. In [48],
the researcher stated that the various Opportunity in human life
made an opinion leader. They proposed that if a user received
first-hand information from recognized sources, then no need to
rely on any other social media. The gathered information itself
sufficient to make them opinion leaders. In [49], the authors de-
fined the role and explored the influence of opinion leaders on the
diffusion process of famous mobile games. They experimented
that if an opinion leader promoted the game-related topic, the
diffusion rate of those words is relatively too high as compared
to promoted by the user-generated content. In [41], the author
proposed a unique K-means framework includes three phases;
the first phase identified the list of opinion leaders having a max-
imum local reputation, the second phase espoused an innovative
active game model to discover the locally Pareto-optimal com-
munity structure and the last stage used the opinion dynamics
model for creating the growth of the sentiment matrix. One of
the applications of the opinion leader in Smart Grid that consists
of ‘Prosumer’ who not only uses the energy but also generates the
energy used by other grids or consumers. In [50], the researcher
addressed the new approach that used the individual prosumer’s
energy density to find leader prosumers in Smart Grid. They
also discovered a dynamic game model to identify the prosumer-
community group structure and experimented on various Smart
Grid datasets to validate the performance of the system.
In a nutshell, every theory provided its way to the imple-
mentation of opinion leader identification. Hence, we found the
following shortcoming in the existing approaches and also ac-
knowledged our involvement in solving the limitations.
3
• Deep convolution: The structure of the mentioned ap-
proaches is very convolution and challenging to
understand—most of the algorithms, statistical equations,
procedures, models, and frameworks merely understand-
able. We follow the underlying and uncomplicated mecha-
nism for the research. Both the proposed algorithms are ap-
parent, easy to understand, and include a rational formation
for opinion leader detection.
• Absence of human reasoning: The brighter side of the
social network is human reasoning, but unfortunately, no
approach made a serious effort in this area. Although some
strategies focused on the user’s characteristics related to
social status, knowledge, skills, and the total number of con-
nections, in this paper, we used the game theory approach
to represent human reasoning up to some extent. We also
used trust, an essential factor, for making a strong relation
in the social network.
• Small datasets: the selection of appropriate datasets is ex-
ceptionally supportive of the effectiveness of the procedure.
As we have observed, most of the other approaches chosen
online blogs, community, forums, and review communi-
ties for their experiment, evaluation, and outcomes. Only
limited procedures utilized the real social network dataset
for evaluation purposes. In this paper, we have used the
two real datasets for experimental analysis; yet the pro-
posed approach sustains all kins of datasets for evaluation
purposes.
• Limited solutions: Nowadays, a single solution is not so
much appropriate for controlling all the situations. Similarly,
in the social network, a particular answer is insufficient to
find the opinion leader. In the earlier studies, researchers
provided an explanation for the specific scenario with some
limitations. Here, we have provided the four alternative
solutions to receive the payoff by the player in each strategic
move. We also used the Shapley value to compute the
expected average marginal payoff for every user needed to
make a potential coalition.
• No rational decision making: Most of the approaches used
conventional methods and functions based on graph theory
and complex network to identify the opinion leader. There
is a lack of rational decision making to make logical and
sensible decisions. The main benefits of rational decision
making are to choose the best possible impartial solution
over alternatives. We proposed a unique amalgamation of
social network analysis and game theory to produce the best
possible solution.
• No concept of the synergic coalition: In the synergetic
coalition, various users participated in a group and con-
veying a compensation obtained from their marginal pay-
off and guaranteed that a consistent and robust coalition
formed. Till now, no approach considered this kind of group
formation to find the opinion leader. The first time we
have proposed such type of composition for opinion leader
detection.
• Leverage of Data mining techniques: Data mining tech-
niques are beneficial to draw fruitful outcomes from the
unrefined data undoubtedly. Classification, clustering, text
mining, outlier analysis, and many more techniques used
by different analysts to detect the opinion leader in the
social network. But these methods are not so practical for
social network analysis due to the high volume of the users
and the active environment of the social network. Although,
these procedures are very much industrious and productive
for statistical analysis in health, marketing, financial bank-
ing, pattern visualization, and many other applications [51,
52]. We used a different approach that uses the statisti-
cal measurement, arithmetic calculations, and game-theory
based policy to identify the opinion leaders.
4 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158

Table 1
Summarized review of literature papers.
References Dataset Mechanism Benefit Limitations
[23] Online forum Text mining and social network analysis based new Able to extract opinion Limited text database;
method have suggested in which positive and negative automatically; useful for Only applicable for a
opinions formed. Opinion leaders selected based on their business modeling static data set
total positive opinions.
[24] Reviews of Apple’s Proposed a model in which the connections among the Simple and easy to Applicable only for online
iPhone user’s cluster detected. Social behavior of the users implement; Helpful in communities; Lack of
analyzed based on these connections. Finally, the text recommendation system validation database.
mining technique used to identify opinion leaders. and trend analysis.
[25] Epinions dataset An extensive Advogato trust metric proposed that Manage to get to Trust metric is not
encourages the recognition of reliable user concerns with authorization in the valuable for worldwide
other users. Next, recursively diffuse the limit of network; provide a express connections;
designated users all through their system. Finally, it solution for the Access only the binary
exhibits the capacity-first maximum flow technique for recommended system. relations in the network.
finding the most grounded way appropriate for finding a
set of valid users.
[26] Synthesized dataset Explored the opinion leader based on the best-promoting Useful for marketing and Adopters information
decision, dissemination speed, and the most extreme total business purpose; needed; Limited
number of adopters utilizing age model. Besides, the Effective for indirect centralities used for
opinion leaders with high sociality are considered for recommendations calculations; Viable
quick dissemination based on the reenactment characteristics of the
result-users with high disjointing centrality considered as products neglected.
the most extreme aggregate number of adopters.
[27] Educational blogs A mixed approach proposed contained user connections, Very helpful in online Covered only a few
dataset, user behavior, user prominence, and time of responding forums and blogs; high online forums and blogs;
Parent–child forum recognized. The opinion leader discovered based on four execution as far as Only topic-specific
dataset features. centrality. constraints considered.
[4] Amazon.com A model is designed to detect domain-sensitive opinion Helpful for product Limited scope; The user’s
dataset leaders in online forums and networks using a model. For dissemination; Suitable interest needed for
the ranking purpose, identify the customer’s domain of for inclusion and evaluation purposes.
interest, user’s specialty, and transition probability matrix. authority.
[28] Facebook dataset, Characterized an exponential time-decay formula to The proposed model is Too complicated to
citation network gauge the impact of leaders and build the chains of helpful to find influence understand; not suitable
dataset (Google leaders’ action specific impact. At that point, next, path; High accuracy and for all datasets; Lack of
scholar & DBLP) manufacture the global chains by normalizing every precision. parameters to chose the
conceivable user’s activities. path.
[29] Web-based Chinese User’s activities measured based on the text submitted on Useful for marketing and Only compared the
stock forum the board. Next, applied the clustering algorithms on the business strategies; Easy outcomes with PageRank
dataset user’s information and set of appropriate opinion leaders to implement. centralities; Utilized only
identified. Further, the critical opinion leader discovered for online blogs and
based on the sentiment analysis to find the real price forums.
movement patterns.
[30] Synthesized dataset The proposed algorithm detects the central issue in the Better average precision Do not consider the total
domain and then measure the leadership score of each and precision–recall; Easy number of retweets; The
user in the area. The opinion leader discovered based on to understand output is topic
the highest leadership score. dependent.
[31] Douban.com An algorithm proposed to identify the node status based Converged in lesser Only page rank used for
on importance matrix signaling spreading, i.e., how a iterations; High accuracy compared results; Few
node penetrated the entire network. The opinion leader and efficiency. features used to manage
selected based on the node status and importance. the importance matrix.
[32] Epinions dataset An innovative framework depends on trust in which total Better in terms of Less accuracy; Uses
trust value (TTV) and opinion leader selected found on in-degree, out-degree; limited trust metrics.
maximum TTV. Hybrid IO-degree method,
[33] Mobile01 forum An efficient procedure, D_OLMiner, suggested for Highly efficient; Compared the outcomes
identification of opinion leaders in a robust social Implemented dynamic with few metrics;
network. Next, a network rising technique proposed to social network; Better Applicable only for online
build a dynamic social network to recognize the network impact flooding forums
community, and unravel the impact covering the issue
and decrease the total calculation time.
[34] Mobile01 forum Addressed the innovative OLMiner algorithm used the Overcome overlapping Not suitable for all social
behavior and standard neighbor connection to reduce the influence problem; High network datasets; Limited
total number of generated candidate and choose the list efficiency and scalability. similarity metrics used.
of opinion leaders.
[35] turnbackhoax.id Implemented two approaches: Edge weighting and Suitable for rumor Limited centrality
Centrality weighting; Edge weighting used to find the spreading in the measures and
opinion leader by counting the total number of tweets network; identify the relationship used;
and retweets. Centrality weighting is used to assigned importance of each edge; Applicable only on a few
weight to each centrality for finding the opinion leader Highly accurate. social networks.
more accurately.

(continued on next page)

 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158 5

Table 1 (continued).
References Dataset Mechanism Benefit Limitations
[36] KOL dataset using Opinion leader identification is considering a multimodal Better subjective and Highly multifaceted;
Instagram API, skilled job. An innovative subspace learning algorithm quantitative outcome; straight-direct
Synthesized dataset addressed to discover the low-dimensional discriminator Presented multimodality modification is not
from social media data to select key opinion leaders. and high-dimensionality sufficient for the whole
characteristics. informational collections.
[37] Synthesized dataset Opinion leader detection dependent on the new closeness Progressively productive Highly complex; Only
based calculation that incorporates a distinctive kind of and precise; designed implemented on a static
collaboration among the users. The calculation updated independent dataset.
additionally relies upon the communication time and cascade model; show
postponement of nodes in the system. more power than
betweenness centrality.
[38] Sina micro-blog The opinion leader’s selection is based on topic-sensitivity Profoundly effective; Pre network structure
dataset constrained. At first, the spread properties and the user’s Improved Independent and node information
topic pertinence qualities estimated and haphazardly Cascade model; the lesser required; Due to the
accept 10% of the user. Opinion leaders classification number of seed nodes assorted variety of topics,
dependent on the time of data diffused and the number needed with limited data not appropriate to work
of tainted nodes. spreading time. with a variety of datasets.
[39] Synthesized dataset Proposed a methodology dependent on the trust age Better outcomes as for Utilized static informal
model and fluffy rationale. In the first place, fuzzy exactness and precision; dataset; More rules and
standards assessed to discover trust esteems. Positioned implementing and parameters needed to
the Opinion leaders dependent on the highest evaluating human produce trust accurately.
prominence value. reasoning.
[40] Chinese Sina BBS A PageRank based algorithm called Hybridrank proposed High accuracy over Only considered link,
considering topic sensitivity and temporal features. Topis PageRank algorithm; Easy topic and temporal
sensitive analysis used to find groups while temporal to implement characteristics; applicable
feature analysis used to obtain the influence of opinion for a small dataset
leaders over time.
[41] Facebook, Twitter, Proposed a framework based on a discrete-time dynamic High Average cluster Complex in nature; Not
Google+ model. Opinion leaders identified based on densely quality; More accurate suitable for overlapping
connected components having a similar opinion vector. and robust opinion communities.
dynamic model.
[42] Twitter Defined a unique component milestone centrality Proposed innovative Experiments performed
includes interest and exclusivity on some topics by centrality measures; Easy only on few topics;
participating users. The users with the highest milestone to understand; No need Milestones needed to
centrality elected as opinion leaders. to review relationship. validate the results.
[43] Local motor Proposed a method included degree, median, and near Robust; Straight-forward Lack of the degree of
community centrality as multi-features attributes. The benefit and easy to understand. promotion; Only used for
indicator and cost indicator used to improve the accuracy static datasets.
of the method. Opinion leaders are selected based on the
highest multi-features characteristics.
[44] Slashdot dataset Addressed a swarm intelligence based firefly algorithm Simple and easy to The static dataset used
for uncovering the local and universal opinion leader implementable; Improved for implementation; few
communities. First, an improved Louvain community accuracy and precision. characteristics acquired
detection procedure addressed, and users elected based for attractiveness
on their attractiveness value. calculation.
[45] Travel community Opinion leaders identified in virtual travel communities Used SNA in virtual Only applicable for few
dataset based on construal, content influence, and action to travel community; datasets; Consider few
measure the influence on consumer’s decisions. measure opinion leader parameters for
influence efficiently identification.
[46] Wiki-vote and Suggested an algorithm (SNWOA) to locate the opinion Improved whale Compared the results
synthesized leader using estimating the notoriety of the users and optimization algorithm with few meta-heuristic
datasets different standard enhancement functions. At first, they used; Better performance algorithms; The Static
figured the target function by utilizing the user’s for large datasets. dataset used.
centralities and conveyed the proposed calculation using
various streamlining functions to discover all kinds of
opinion leaders.
[47] Higgs Boson data Proposed an approach based on in-degree and out-degree Simple and easy to Only a few centralities
from Twitter to identify the opinion leader. Next, rank the user by implement; Less complex. considered; Limited
their BC value. It also implemented the standard Louvain scope; Used static
method for community detection. dataset.

3. Preliminary 3.1. Social network analysis

A social network considered as a directed graph in which the

In this section, we presented the underlying phenomenon and
trends related to the social network and game theory needed
for obtaining in-depth knowledge about the matter. Initially, we
discussed some factors related to social network analysis and
later associated with game theory and Shapley value in detail.
node denotes the actor while the link between the actors indi-
cates the relationship between them. If a graph consists of n num-
ber of nodes, the relationship can be depicted using n x n matrix.
In the directed network, a weighted value associated with each
node, and if the value is non-zero, a weighted relationship exists
between two nodes i and j. The considerable upper value beyond
6 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
the standard value indicates the significances of the nodes in the
link. In the social network, trust is the leading factor that plays a
prominent role in establishing a relationship [53]. Trust is excep-
tionally supportive to determine the degree of strangeness within
the network [54]. There are various factors such as clustering
coefficient, density, network size, indegree–outdegree covariance,
transitivity, homophily, etc. affect the degree of connection in
the network [55–57]. Now we discussed the underlying social
network concept in brief as follow:
• Indegree: Defines the number of a user connected to the
user.
• Outdegree: Defines the number of users connected from a
user.
• Indegree–outdegree covariance: It defines the covariance
relation between the in-degree and outdegree of the nodes.
• Clustering coefficient: The degree to which nodes are
tended to make a cluster known as The clustering coefficient
of a network.
• Network density: The network density defined as the ratio
between the total number of existing edges and the entire
possible number of edges in the network.
• Transitivity: The concept of transitivity based on triad in
which if a node A is the friend of node B, node B is the friend
of node C, then node A is also a friend of node C. In the real
world, sometimes, the same concept may not work in some
situations where a friend of a friend is not a friend.
• Homophily: The same types of nodes, based on their at-
tributes, may likely to attach or attempted to make a union
as compared to different nodes in the network [58].
• Network size: The total numbers of nodes in the network
define the size of the network.
• Closeness centrality: Average distance of the shortened
path between the node and all other nodes in the network
specifies the closeness centrality of a node in the network.
• Betweenness centrality: The ratio between the total num-
ber shortest path passes through the node, and the total
number of the shortest path between every node in the
network signifies the betweenness centrality of a node. Be-
tweenness centrality means the power of the node in the
network.
• Degree Centrality: The total number of connections that a
node has in the network, i.e., the summation of the total
number of ties sent and received to and from the node
indicates the degree centrality of a node in the network.
• Eigenvector Centrality: The Eigenvector Centrality of the
node depicts the authority of a node on its associated neigh-
bors in the network. It is calculated using Eigenvectors and
concerns Eigenvalues. The particular case of Eigenvector
centrality is PageRank centrality determines the transitive
impact or availability of nodes in the network.
3.2. Game theory
Game theory offers a prescribed, structured, organized, ana-
lyzed, and systematically evaluated scenario that includes some
set of players and a set of rules followed by each player [59].
Players play the game according to the rules. Whenever the player
chooses a strategy and performs an action, an outcome produces.
There is a payoff associated with each strategy stroke and can
be of the monetary or non-monetary type, i.e., in the form of
happiness, joy, or completeness. There is also uncertainty in the
outcome, i.e., we cannot forecast the result of the strategy in
advance. There are also some predetermined rules to play a game,
but either the player may change its plan, or the environment and
conditions may change results insufficient desired payoff. An ap-
proach is called optimum if it increases the player’s payoff in each
Fig. 2. The trustor–trustee tree structure for extensive form game.
move. So, some strategies move are exquisite and produce higher
payoff. Game theory also follows the no-cheating phenomenon,
i.e., no player can deceive other players in the game. So, it depicts
the strategies that show how a player can win the game without
any cheating [60,61]. Game theory classified the game into two
categories: Normal form game and Extensive form game.
3.2.1. Normal form game
In normal form game, there are a finite set of players, and each
player has a limited set of strategies. There is also pay off function
that assigned some pay off to each player depends on the player’s
strategy and other’s player strategy. For example, in the Prisoner’s
Dilemma, there are two persons with their plans cooperate(C) and
defeat (D). If any of the people change their policy, both lose their
optimal payoff [62].
3.2.2. Extensive form game
An extensive form game, a well-formed tree is defined in
which every player has the information about the other players
like their strategies, possible outcomes of a sequence of moves,
payoff associated with all game results. The tree structure in-
dicates the graphical representation of a player’s decision at a
specific moment. The structure of the tree includes a root node
that represents nature to move and responsible for the starting
of the game. Each leaf node represents the n-tuple payoff vector
that includes the payoff associated with each player obtained at
the end of the moves. Each node of the tree indicates the possible
level of the game from which the player chooses its strategy for
achieving the next level. Each edge of the tree depicts the possible
action taken by the player at each level. Once a user reaches a
particular level, there is a probability distribution function over
an edge used by the player to reach another stage. The terminals
nodes of the tree contain the final payoff associated with every
player involved in the game and also exemplify the ending of the
game [63,64].
For example, considered two players A and B are playing with
two strategies X and Y in Fig. 2. At the ending of the game, both
players receive some payoff. If player A chooses strategy Y and
player B adopts strategy X, the payoff might be one for player
A and two for player B. Similarly, if both the payer chooses the
strategy Y, both get the payoff zero.
3.2.3. Shapley value
As we discussed earlier that ‘Trust’ is a prominent feature
that affects the degree relationship and makes a strong or weak
connection between the users. The trust between the users also
concerned with the mentioned characteristics. For calculating the
trust among the users, we use the concept of Shapley value [65].
The Shapley value measures the expected payoff obtained by the
player in the coalition game theory. The Shapley value represents
the contribution of the player in a coalition having n number of
players [66]. In other words, we can say that if the group of people
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
participated in a coalition and got some payoff, how the payoff
divided among the players fairly. The main advantage of Shapley’s
value is to provide the expected marginal payoff based on the
participation of the player in a game [67,68]. There are various
ways to calculate the Shapley value with some constraints [69].
In our paper, we make a hypothesis that either all the players
belong to the same coalition or are nearby to at-least x-neighbors
who are in the coalition. The main reason behind this assumption
is that in the real world also, users generally interact only with
those persons who follow the same behavior as by them and
likely to make a group with them. In this case, a game can be
represented as (N, v ) where N represents the total number of
players in the game, i.e., N = {1, 2, 3. . . n} while v indicates the
n
function v ϵ R2 −1 served using Eq. (1).
{ }
0 if C = 0
v=
{v ∈ C ||N(v) ∩ C| ≥ x} else
It is evident that if deg(node) < x, all the node belongs to the
same coalition otherwise divide the network into other coalitions.
So this is required that the deg(node) ≥ (x + 1) for more precise
calculation. In the coalition game, every player has its valuable
reserve, and once the player alliances in the game, it may produce
some collective synergy that is higher or may be lower than the
sum of their resource at time t. For example, in a game, if two
players I and j have individual payoff is 3 and 2 respectively. Still,
if both players collate in the game, the collective synergy might be
six (multiplicative factor), which is higher than the sum of their
payoff [70]. Therefore, Shapley value analyzed the contribution
of each user once they come together and make a cluster. The
Shapley value of an individual based on payoff can be represented
using Eq. (2).
N−i
∑ |C|! (n − |C| − 1)!
SP(v) = {v(C ∪ i) − v(C)}
n!
Cϵ 1
In the above equation {v (C ∪ i) − v (C )} represents the marginal
payoff received by the user in a collation game C. Hence, the
above equation depicts weight to the expected contribution.
4. Proposed methodology
In the social network, it is not a straightforward task to iden-
tify the opinion leader; still, we endeavored to find the opinion
leader in the social network. As we have discussed earlier, that
game theory categorized into two forms. For implementing the
proposed approach, we used the extensive game theory form in
which a trustor–trustee tree generated based on trust. The trust
and centrality measures are used as an attribute to define the
user characteristics in the network [71]. We hypothesized that
each user behaves like a player in the network in which trust and
other centrality measures, as we discussed above, considered as
attributes help to find the marginal contribution of a user in the
game.
Initially, we measured the degree of trust based on the user’s
frequent interaction patterns. Trust can be unidirectional or bi-
directional depends on user behaviors. If a person has some prior
information or recommendation about past experiences and per-
formance, it increases the degree of trust in another person [54,
72]. As shown in Fig. 3, trust is classified as direct, indirect, and
recommended. In direct trust (DT), a user trusts directly another
person without any influence. In contrast, in the case of indirect
trust (IDT), a user trusts another person indirectly by the impact
of some other person. In the case of recommendation trust (RT),
trust derived from the recommendation of other people based on
their experiences. The trust degree Txy is indicated in Eq. (3).
Txy = , ,
f(dxy dyx dzxy , dzyx , rx , ry )
7
In the above equation, dxy and dyx shows the mutual degree of
trust between the user x and user y, dzxy and dzyx indicates the
trust degree of trust through user z, while rx and ry signify the
user’s reputation.
In this paper, we have considered trust as a significant factor
in establishing a relation between the users. We have considered
the three psychological collective elements Goodwill, Power, and
Uprightness, for measurement of the user’s fidelity. These three
elements reflected the aggregate trustworthiness of the user and
also replicated the willingness of other users to establish a friend-
ship. The pseudo-code for the trust score calculation shown in
Algorithm 1.
In the above algorithm, we have used the three constant
threshold variables k, l, and s such that (k, l, s) ϵ R and varies in
the interval [0.5, 1) for all three types of trust, respectively. The
(1) threshold value of the entire three variables is network-specific
and depends on network structure and dynamics.
Once we have computed the trust, the next task is to calculate
the Shapley value in a coalition. In a network, Users made the
coalition based on their shared interest, dyadic and triadic clo-
sure, strong and weak ties, and many other factors based on the
nature of the network [73,74]. In this paper, we have proposed a
new approach in which we considered all possible combinations
of the coalition of users. If the synergy of marginal contribution
of the entire users in a coalition is higher than the amount of
marginal contribution of users of any other group, all the users
belong to that coalition considered as the opinion leaders in the
network. In the case of the same synergy, the coalition has the
users with higher Eigenvector centrality considered for selection.
The proposed approach is unique and has lots of potentials to
present different variations based on different parameters.
In the real world, it happens that some of the players in the
game have no much experience and cannot portray the game
(2)
accurately. Most of the time, they betrayed by their peers and
colleagues. An inexperienced player may make decisions based on
the cunning movement made by other players. In some situations,
there is a probability that the other player may also be influenced
by the other player’s strategy up to some extent and modify their
actions according to them. Thus, considering the above facts, we
have considered the following four alternatives.
Solution 1: In this case both the player tries to persuade another
player to accept or agree on their plan or change their strategy
according to the plan of another player. Of course, each player
is not easily accepting the strategy of another player; therefore,
there is a fixed payoff, called p, associated with each player which
is defined as follow:
p = +x if the other player changed their plan
p = −y if the other player unchanged their plan
p = −x if the player changed their plan
p = +y if the player unchanged their plan
Thus, the payoff table for both the player A and B is as follow:
Payoff for player A
B unchanged plan B changed plan
A unchanged plan 0 +y + x
A changed plan −x − y 0
Payoff for player B
B unchanged plan B changed plan
A unchanged plan 0 −y − x
A changed plan +x + y 0
Therefore, the payoff matrices for both the player A and B are
as follow:
[ ] [ ]
0 +x + y 0 −x − y
MA = MB =
(3) −x − y 0 +x + y 0
8 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158

Fig. 3. Trust classification models.

In the above case, we can observe that if the player A unchanged provide another optimal solution based on centrality associated
their plan, player B have two option either change or keep their with the intermediate solution.
plan. In both cases, the payoff would be –x–y and 0, respectively. Solution 3: In this solution, we have added a new parameter
So this is the best option to choose 0 instead of –x–y, i.e., another called ‘distance’, say d, which signifies the mean centrality be-
player unchanged their plan. The above discussion concludes that tween the users. In the social network, centrality plays a crucial
both the player would not get any payoff and keep their plan role in defining the importance of a node. The centrality of the
unchanged and never reaching on an agreement. Thus, we have user depends on the network dynamics and network structure. As
provided another solution in which both the player agreed on an the dynamics of the network changes, the centrality of the user
intermediate solution.
also varies over time. It is evident that if the centrality of a user is
Solution 2: In this solution, we suggested an intermediate solu-
far above the ground, a user is reachable and accessible by most of
tion called ‘agreement’, i.e., both the players do not change their
the other users easily and may influence them with their behavior
strategies and consent on an intermediate solution. In this case,
we introduced the fixed payoff p for both the player as follow: and technical knowledge. In the proposed approach, we have
p = +I if the other player modify their plan towards the inter- used the Betweenness Centrality (BC), Closeness Centrality (CC),
mediate solution Degree Centrality (DC), and Clustering coefficient Ci to define the
p = −I if the player modify their plan towards the intermediate distance dij between the users i and user j in the network using
solution the Eq. (4).
Thus, the payoff table for both the player A and B is as follow: n √
∑ BCi ∗CCi
dij =
2
+ λCi + ρ Cj (4)
Payoff for player A DCi
iϵ N −j
B unchanged B changed B agree
plan plan In the above equation, both λ and ρ are the weighted coefficient
A unchanged plan 0 +y + x +y + i and used for balancing the distance d. the value of λ and ρ
A changed plan −x − y 0 −x + i between [0, 0.5]. We also found that, if the clustering coefficient
A agree −y − i +x − i 0 of the user increases gradually, the distance between the users
also increases. Now we proposed the payoff p for both the player
Payoff for player B A and B are as follow:
B unchanged B changed B agree p = +x if the other player changed their plan
plan plan p = −y if the other player unchanged their plan
A unchanged plan 0 −y − x −y − i p = −x if the player changed their plan
A changed plan +x + y 0 +x − i p = +y if the player unchanged their plan
A agree +y + i −x + i 0 p = +i+ 1d if the other player modify their plan towards the
intermediate solution
Therefore, the payoff matrices for both the player A and B are p = −i+ 1d if the player modify their plan towards the intermedi-
as follow: ate solution
0 +x + y +i + y
[ ]
Thus, the payoff table for both the player A and B is as follow:
MA = −x − y 0 +i − x
−i − y −i + x 0 Payoff for player A
0 −x − y −i − y B unchanged B changed B agree
[ ]
MB = +x + y 0 −i + x plan plan
+i + y +i − x 0 A unchanged plan 0 +y + x +y + i+ 1d
Again, in the case, we observed that if the player unchanged their A changed plan −x − y 0 −x + i+ 1d
plan or agree on an intermediate solution, Player B has three
A agree −y − i+ 1d +x − i+ 1d 2
alternatives, the first one is to change the plan, the second one d

is unchanged, and the thirds one is to agree on an intermediate

solution. In all the tree alternatives, the payoff for player B would Payoff for player B
be –y–x, 0, and –y–i, respectively. So the best option for player B B unchanged B changed B agree
is to choose payoff 0 and unchanged their strategy. We analyzed plan plan
that the best option for both the player is to unchanged their A unchanged plan 0 −y − x −y − i+ 1d
strategy and did not receive any payoff. Thus, in this case, also
both the player never ended with a proper solution and needed
A changed plan +x + y 0 +x − i+ 1d
some better solution to receive the optimal payoff. Thus, we A agree +y + i+ 1d −x + i+ 1d 2
d
L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158 9
10 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158

Fig. 4. A synergetic coalition of two networks.

Therefore, the payoff matrices for both the player A and B are Hence, the payoff matrices for both the player A and B are as
as follow: follow:
⎡ ⎤ ⎡ ⎤
1 1
0 +x + y +i + y + ⎥ ⎢y((1 − ub ) − (1 − ua )) y(1 − ub ) + xua y(1 − ub ) + i + ⎥
⎢ d⎥ ⎢
⎢ d⎥
⎥
⎢
1⎥ MA = ⎢ −xu − y(1 − u )
⎢ 1 ⎥
MA = ⎢ −x − y
⎢
0 +i − x + ⎥ b a x(ua − ub ) −xub + i + ⎥
⎢ d⎥
⎢
⎢ d ⎥
⎥
1 1 2 1 1 2
⎣ ⎦ ⎣ ⎦
−i − y + −i + x + −y(1 − ua ) − i + xua − i +
d d d d d d
⎡ ⎤
1
⎡ ⎤
1 ⎢y ((1 − ub ) − (1 − ua )) −xub –y (1 − ua ) −y (1 − ua ) − i +
0 −x − y −i − y + ⎥ d⎥
⎥
d⎥
⎢ ⎢
⎢ ⎥
⎢
1⎥ MB = ⎢
⎢ 1 ⎥
M B = ⎢ +x + y ⎢ y (1 − ub ) + xua x (ua − ub ) xua − i +
⎢ ⎥
0 −i + x + ⎥ d ⎥
⎢ d⎥ ⎢ ⎥
⎣ 1 1 2 ⎦
1 1 2 y (1 − ub ) + i +
⎣ ⎦
−xub + i +
+i + y + +i − x + d d d
d d d
In this case, we found that if both the players do not change
In this case, we analyzed that if both the player agreed on the their plan, still both the player may get some payoff based on
intermediate solution, the payoff for the player is 2d . Although the parameters ua and ub with the constraints that 1 > ub > 0, 1
this solution produced better results as compared to solution 3, > ua > 0. Similarly, if both the player agreed on the intermediate
again, the same problem occurred in which if any of the players solution, the payoff may remain the same for all. Therefore the
unchanged their plan, another player may have three choices proposed solution motivates the player to plan their strategy
−y−x, 0, and −y−i+ 1d to take the payoff. So the best option for and choose the next step accordingly. Suppose, at time t, both
another player is to choose payoff 0 if (y+i) > 2d and agree on an the player takes the decision based on the previous progress,
intermediate solution. In the particular case, another player may the time progression needed to convince another player can be
receive the payoff 0, if both the players changed or unchanged represented using Eq. (5)
their plan simultaneously. So again, for overcoming this problem,
we suggested a new solution along with the parameter d. EA (t + 1) = EA (t ) + µ(EB (t ) − EA (t )) if 0 < µ < 0.5 (5a)
Solution 4: In the real-world, it has been observed that it is not EB (t + 1) = EB (t ) + η(EB (t ) − EA (t )) if 0 < η < 0.5 (5b)
necessary that each time a user completely disagrees with the
other player strategy. There is a probability that up to what level In the above equation, both µ and η both are the configurable
a player convinces another player with their plan. In this solution, variables that lie between [0, 0.5]. Finally, we noticed that if ua
again, we added a new parameter called ‘u’, which signifies the < (i + 1d ), the player A decides to take the decision ‘agreement’
degree of inducement in the game. We define two probability ua whatever the strategy is chosen by the player B. Similarly, if ub
and ub that signifies the probability that up to what level, player < (i + 1d ), player B chooses the decision agreement, whatever the
A influences player B, and player B, respectively influence player plan decided by player A.
A. So, in this game, the payoff of each player is influenced by two
parameters ua and ub . If player A is not able to influence player 4.1. Calculation of synergy
B, then the (1- ua ) be the probability to receive the payoff. Now
we discuss the payoff for both the player as follow: A synergy of a network is defined as the combination of
ability or power of nodes that produces more value than the
Payoff for player A separately two or more nodes can do. A regular machine can do
B unchanged plan B changed plan B agree more productive work as compared to another machine, if all
A unchanged y((1- ub )-(1- ua )) y(1- ub ) +xua y(1- the components of the machine coordinate seamlessly with any
plan ub )+i+ 1d mutual inference [70]. Likewise, in social networks, a coalition
A changed −xub −y(1-ua ) x(ua - ub ) −xub +i+ 1d of a lesser number of organized people might produce more
plan productive and innovative outcomes as compared to a group of
A agree −y(1-ua )−i+ 1d xua −i+ 1d 2 more number of disorganized people. The main phenomenon
d
of synergy is based on structural holes in the network [75].
A structural hole created when a single node connected with
Payoff for player B
the vast network, and the same node also joined with another
B unchanged plan B changed plan B agree
node related to an extensive network. So these two nodes are
A unchanged y((1- ub )-(1- ua )) −xub −y(1-ua ) −y(1-
the only medium through which information can pass from one
plan ua )−i+ 1d
group to another group. These nodes are responsible for shar-
A changed y(1- ub ) +xua x(ua - ub ) xua −i+ 1d ing knowledge and information [76]. Researchers suggested that
plan these nodes having a high impact on network operations because
A agree y(1- ub )+i+ 1d −xub +i+ 1d 2
d
they may affect multiple links simultaneously. In Fig. 4, consider
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
Fig. 5. Flow chart of the proposed algorithm.
the two networks having a node with higher degree centrality
(blue color node), and most of the other nodes connect with that
node. If any node of a network needs to share information with
11
another network, all the data might pass via this node, so this
node plays a critical role in the synergetic coalition of networks.
Two or more nodes are likely to be integrated when the merg-
ing outcome is extremely higher than the expected outcomes. For
measuring the synergy of the merger, we used Eigen centrality
(EC) as the major component as the source node would attempt
to merge with only those target nodes having a higher Shapley
value. So the synergy ωij of the merger of two nodes and total
synergy of a coalition can be measured using Eq. (6).
1{ } 1
ωij = (EC)i + (EC)j ∗ ∂{(SP)i + (SP)j } (6a)
2 2
x x ( ω )c
ij
∑∑
ϕ= δ∗ (6b)
x
i=1 j=i+1
where x is the total number of nodes in the coalition, c is the
specific condition in which the user behaves, δ , and ∂ is the
coordination rate defines the coordination among the users. The
value of c, δ and ∂ varies between 0 and 1. The flow chart of the
proposed approach shown in Fig. 5, and pseudo-code revealed in
Algorithm 2.
4.2. Complexity of the algorithm
The complexity of the algorithm determines the total amount
of time and memory needed for the implementation of the en-
tire program for specified input. Here, we have anticipated two
unique algorithms; the former one is addressed to calculate the
degree of trust, and the later depicted the identification of the
coalition in the network. In the first algorithm, we measured
the direct, Indirect, and Recommended trust based on the fun-
damental behavioral attributes of goodwill. For measuring the
direct trust, all nodes are considered directly connected with the
source node. Hence the complexity of the direct trust is O(E).
Next, for measuring the indirect trust, we used the depth-first
search along with measuring the clustering coefficient of the
node. Thus, the time complexity of indirect trust measurement is
O(V + E). In the case of a recommended trust, we used the node’s
adjacency matrices to measure the recommendations suggested
by its neighbors. So, the complexity of the recommended trust is
O(V + E). Therefore, the overall time complexity of the algorithm
is O(V + E) + O(E) ≈ O(V + E).
In the case of game theory-based opinion leader detection,
initially, we have to calculate the conditional probability of the
nodes in the network based on Bayes’ theorem. In this case, all
the users are independent of each other; the action or decision
taken by them is to depend on their acquaintance’s actions.
Therefore the time complexity of this procedure is O(V). Next,
we calculated the Shapley value of each user in the coalition.
The complexity of Shapley value calculation depends upon the
type of input, i.e., how the input given in the problem domain.
As we discussed earlier, we have assumed that either the user
belonging to the same coalition or are contiguous to as a min-
imum x-neighbors who are in the coalition. Therefore, the time
complexity to evaluate the Shapley value of each node is O(|V | +
|E |).
Further, we have performed a calculation on the whole pos-
sible number of permuted coalitions to measure the probability
that which permutation produces the highest synergy. We used
the Shapley value and Eigenvector centrality to measure the
synergy of each coalition. Thus, the time complexity to measure
the synergy for each coalition is O(|V|2 ). Next, we ranked all the
coalitions based on their synergy. The complexity of this step is
based on the total number of coalitions, say e, identified in the
network. So the time complexity of this step is O(e). Thus, the
12 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158

overall time complexity of the proposed algorithm is O(|V | + |E |) Table 2

+ O(|V|2 ) + O(e) ≈ O(|V|2 ) Comparison of time complexity between the proposed algorithm and standard
baseline centralities.
Generally, centrality measures are considered as baseline
S No Approach Time complexity
methods to find the influencer node in the large scale social
network, as we have discussed in Section 3.1. So researchers 1 Degree centrality O(V2 )
2 Closeness centrality O(V3 )
made an effort to exploit various centrality variations in their
3 Betweenness centrality O(VE + V2 )
research. Now, we compared the time complexity of the proposed 4 Eigenvector centrality O(V3 )
algorithm with the standard baseline centralities in Table 2. 5 PageRank centrality O(V3 )
6 Proposed O(V2 )

5. Experimental results

For implementing the suggested algorithm, we used the two
real networks; Wiki-vote and Bitcoin OTC trust weighted signed
dataset. As we have mentioned earlier, no conventional method
present to measure the experimental quality of the algorithm
due to the dynamic nature of the network. We used Gephi 0.9.2
software for measuring the network parameters and visualization
purposes [77]. We used python 3.0 and Intel i7 multi-generation
processor to obtain the experimental results [78]. Now, we would
discuss the detailed description of the datasets as follow:
5.1. Dataset
In this section, we conversed on the details of datasets used
for proposed method validation purposes. We used the two
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158 13

real datasets: Wiki-Vote social network and Bitcoin OTC trust Table 3
weighted signed network. Now, we explained the structure and Top-5 coalitions in the Wiki-vote and Bitcoin OTC trust weighted dataset.

configuration of both the datasets. Data set Coalition Total number Collective
opinion leaders synergy
C1 382 278.45932
5.1.1. Wiki-vote social network
C2 478 275.06345
Wikipedia is an open information portal that provides various Wiki-vote dataset C3 684 270.48298
types of information for its users. Wikipedia allows the facility for C4 952 264.95834
its user to add new information and features under the supervi- C5 1689 262.10047
sion of administration. If the user wants to update any technical C1 205 186.99605
details on an issue, a request delivered to the administrators who Bitcoin OTC trust
C2 380 185.26755
have a right to update the information. If a user wants to become C3 538 183.33674
weighted dataset
C4 884 180.38609
an administrator, an application of adminship gave to Wikipedia C5 1087 176.19118
authority. The election of an administrator is either done by the
community users or by the voting mechanism. The Dataset con-
tains all the voting information includes a total number of votes Table 4
Top-10 opinion leaders in Wiki-vote data set.
and a total number of candidates participating in the election. The
Rank Node id Shapley DC CC BC EC PR
node in the dataset indicates the total number of users, while an
value
edge from one user to another user depicts the voting choice of
1 3542 97.53765 0.24915 0.48009 0.32025 0.27582 0.41251
the user. The dataset has a total of 7115 nodes and 103 689 edges. 2 267 97.52764 0.24832 0.46276 0.32586 0.27362 0.41589
The density and average clustering coefficient of the network is 3 1089 97.52116 0.24786 0.48769 0.31254 0.26845 0.42185
0.002 and 0.081, respectively. Some structure holes are valuable 4 5376 97.51986 0.24641 0.46388 0.32817 0.26521 0.40356
to make a bridge between two or more huge clusters [79]. 5 1008 97.51912 0.24465 0.47894 0.32118 0.26559 0.41664
6 387 97.51854 0.24562 0.47106 0.31958 0.26331 0.41985
7 6959 97.51606 0.24378 0.46228 0.32682 0.25487 0.42581
5.1.2. Bitcoin OTC trust weighted signed network 8 2968 97.51594 0.24371 0.46834 0.31584 0.26580 0.41607
Bitcoin OTC trust weighted signed dataset is a trust dependent 9 4822 97.51568 0.24374 0.45732 0.31958 0.25148 0.40931
network in which users trade the Bitcoin using a platform, called 10 3911 97.51561 0.24366 0.46117 0.31705 0.26958 0.41583
Bitcoin OTC (Over-the-Counter) [80]. During the Bitcoin trading,
the identity of the customer is hidden, so it is required to main- Table 5
tain a database of users’ reputation to avoid any counterfeit and Top-10 opinion leaders in Bitcoin OTC trust weighted dataset.
hazard. Such type of network, also known as a who-trust-whom Rank Node id Shapley DC CC BC EC PR
network in which the user grades another user on the scale of value
−10(total distrust) to +10(absolute trust). It formed the web- 1 2829 93.34872 0.32514 0.29201 0.36128 0.22183 0.41006
of-trust network that has a total of 5881 nodes that represent 2 74 93.34773 0.31522 0.29651 0.36982 0.22058 0.41002
the users and a total of 35 593 links that describe the user’s 3 1008 93.34754 0.33201 0.29536 0.36108 0.22015 0.40995
4 3116 93.33543 0.32584 0.28105 0.36471 0.23197 0.41264
trust grading on another user. The density and average clustering 5 1427 93.31008 0.31998 0.27847 0.35927 0.23496 0.40865
coefficient of the network is 0.002 and 0.267, respectively. 6 4275 93.30635 0.30584 0.28853 0.35461 0.22004 0.41106
The main reason to choose these datasets is that the clustering 7 2731 93.30223 0.32115 0.29147 0.36014 0.21986 0.40137
coefficient of both the data set is higher as compared to other 8 694 93.30018 0.31458 0.27651 0.35100 0.21547 0.41042
9 3952 92.29474 0.31025 0.27502 0.34925 0.22010 0.40892
social network datasets. It indicated that the user in the network
10 4937 92.29223 0.32264 0.27984 0.34998 0.21327 0.41072
is likely to make a cluster. The geographical structure, degree
centrality, and adjacency matrix of both the network are shown
in Fig. 6.
user. We have projected the four alternative solutions to calculate
5.2. Analysis and visualization of the experimental result the marginal contribution of each player in a game. Still, the
fourth solution produced better results and took a lesser number
Now, we deployed the proposed algorithm on both the of iterations as compared to other solutions. In Table 3, we have
dataset. In the real world, we found that trust is not considered as mentioned the top-5 coalitions obtained in both the datasets
a discrete value, i.e., not in binary form (yes or no). So in the first along with the total number of opinion leaders with their average
step, we measured the trust among the user based on Algorithm marginal payoff.
1. Initially, we assigned some degree of trust to each user based In the social network, there are also some other SNA centrality
on the user’s experience and knowledge. Further, we designed measures to find the prominence of the node. In Tables 4 and
a trustor–trustee tree by randomly choose a seed user and its 5, we illustrated the Shapley value of top-10 opinion leaders of
subsequent neighbors iteratively. In this process, we measured the coalition having maximum marginal payoff and compare their
three types of trust (Direct, In-direct, and Recommended) de- ranking with other centrality measures.
scribed in the previous section. For measuring the trust, we used In the above table, we can see that the user with the highest
some parameter, say α and β , to obtain the optimal output. The Shapley value does not have the highest degree centralities. Only
value of parameters is decided based on experimental analysis. In a few users hardly maintain the centrality hierarchy, and the rest
Figs. 7 and 8, we found that at α = 0.5, β = 0.75, λ = 0.4, and of the users scattered in the record as per the ranking of the
ρ = 0.4, the proposed algorithm-generated only those opinion Shapley value. Therefore, it does not mean that the node with the
leaders who are significantly influencing the followers. Similarly, higher centralities also has a higher Shapley value. It depends on
for measuring the distance between the users, we used the two the trust, payoff, probability of coalition, and network structure.
weighted coefficients λ and ρ whose value is also identified based During the evaluation, we identified that the solution 4 con-
on experimental analysis. verges earlier as compared to other alternative solutions, as
Next, we measured the conditional probability between the shown in Fig. 9. On the other hand, we can say that solution 4
users using Bayes’ theorem. For implementing the game theory produced the desired outcome in a smaller number of iterations
approach in the network, initially, we supplied 100 units to each and proven its significance to implement the proposed algorithm.
14 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158

Fig. 6. Geographical Structure, adjacency matrix and degree distribution of (a) Wiki-vote network, (b) Bitcoin OTC trust weighted network.

Fig. 7. Parameters value estimation for Wiki-vote dataset.

Thus, the entire obtained coalition and corresponding opinion
leaders evaluated using solution 4.
Further, to justify the superiority of the proposed algorithm,
we compared the experimental results with the SNA measures
based on accuracy, precision, recall, and F1-score [22]. We also
compared the outcomes based on execution time and power
of influence. The identification of opinion leaders depends on
the researcher’s way of experiment analysis and the procedure’s
complexity. On the other hand, we can say that no approach
can classify the opinion leaders accurately until now. The major
problem with the social network is the inaccessibility of opinion
leader’s ground truth. Most of the researchers compared their
findings with the standard SNA measures, and only a few com-
pared the outcomes with the other methods based on some
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158 15

Fig. 8. Parameters value estimation for Bitcoin OTC trust weighted dataset.

Fig. 9. Rate of convergence for all the proposed alternatives to calculate payoff for (a) Wiki-vote network, (b) Bitcoin OTC trust weighted network.

common features in a specific domain. For evaluating these mea-
sures, we needed — True Positive, True Negative, False Positive,
and False Negative [81].
Further, the accuracy, precision, recall, and F1-score can be mea-
sured as:
• Accuracy: Accuracy is the share between total correctly pre-
dicted observations versus the total number of observations.

• True Positive (TP): True Positive includes the estimated TP + TN

Accuracy =
observations identified true by both actual model and pro- TP + TN + FP + FN
posed model.
• True Negative (TN): True Negative includes the estimated • Precision: Precision is the share between the total correc-
observations identified false by both the actual model and tively predicted positive observations versus total predicted
proposed model. positive observations.
• False Positive (FP): False Positive includes the estimated
observations identified false by the proposed model and true TP
Precision =
by actual model. TP + FP
• False Negative (FN): False Negative includes the estimated
observations identified true by the proposed model and false • Recall: Recall is the share between the total correctively
by actual model. predicted positive observations versus the total predicted
16 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
observations by the actual model.
TP
Recall =
TP + FN
• F1-score: F1-score is the amalgamation of precision and
recall, and indicates the harmonic mean of precision and
recall.
Recall ∗ Precision
F1-score = 2∗
Recall + Precision
The compared results based on the mentioned performance met-
rics shown in Fig. 10.
In the above analysis, we can quickly examine that the pro-
posed methodology conferred better results over other SNA mea-
sures for both the datasets. The proposed approach obtained
an approximate 90% accuracy and having 94% precision for the
wiki-vote dataset. Similarly, for the Bitcoin OTC trust weighted
dataset, the accuracy is around 91%, and precision is nearly 94%.
Besides, we also examined the total implementation time needed
to deploy the entire algorithms on the dataset. The whole exe-
cution time includes the preprocessing of the data, computation
of the Shapley value for each user, the formation of the ran-
dom trustor–trustee tree, and measurement of synergy for each
coalition. Generally, the performance of an algorithm depends on
the execution time and the utilization of resources efficiently. As
we discussed the complexity in the previous section, we found
that the suggested approach needed lesser time and about 37%
reduced in the contrast of other SNA measures, as shown in
Fig. 11.
One of the major contributions of the opinion leader is to
diffuse the products over a while [82]. The degree of diffusion
depends on the total number of opinion leaders involved in the
process, the total number of followers, and the time at which
the process took place. The main task in diffusion consists of the
acceptance of any modification or innovation in the product by
the customers in the real world [83]. Sometimes the company
has to struggle many of the years to promote its products in
the commercial market, still not able to convince the consumers.
So, an opinion leader is too supportive of the dissemination
of products and able to resolve the cold start problem in the
recommended system. Moreover, we observed that the proposed
approach required etiquette number of opinion leaders for both
the dataset. At the same time, other SNA measures need com-
paratively more opinion leaders to achieve full adoption by the
customers, as shown in Fig. 12.
We found that the proposed method needed a total of 1859
opinion leaders for the Wiki-vote dataset and 1304 opinion lead-
ers for the Bitcoin OTC trust weighted dataset while the other
SNA measure needed additional opinion leaders for the diffusion
of products and services entirely in the social network.
5.3. Strength and limitations
Each algorithm has some benefits and limitations as compared
to other approaches. In the same way, the proposed algorithm
also has some advantages and boundaries. The leading strengths
and limitations are as follows:
• As we know, the social network structure is dynamic, and
relations between the users change over time [2]. Some
old relations disappeared, and new connections established
between the users. So the duration of a long relationship
depends on the degree of trust over time. Trust is helping to
shape the communication blueprint depicting the individual
expectation on others. In this paper, we have considered
the trust as the main component to gain the maximum
marginal contribution. As the degree of trust is high, the
probability of convincing the user and make a coalition is
also increasing with high synergy. So, we have suggested a
working prototype that represents the real-world scenario.
• We used the game theory approach and made a trustor–
trustee tree that resembles the trust relationship between
the users. We have suggested the four alternative solutions
to obtain payoff based on different parameters. In the exten-
sive game form, each user is unaware of the strategy played
by another player. In the same way, the user in the social
network is not aware of other users but only can trust based
on experience and recommendations.
• We used the concept of Shapley value to identifying the
marginal contribution of a user in a coalition. Shapley value
is used to identify somewhat the kind of expectations that
is needed for forming the synergy.
• Furthermore, the proposed approach gave high performance
over other SNA measures. As we have discussed in the
result section, we achieved approximately 90% accuracy and
94% precision over other baseline centrality measures. The
execution time is also decreased to around 37%, and the total
number of opinion leaders needed for diffusion purposes
also condensed.
• Another benefit of the approach is that as the number of
users in the network regularly extends, the recital of the
model also increases gradually because the accuracy in de-
tecting the coalition also increases with the same effect. The
probability of connecting with the users and the degree of
trust might be measured with more number of observations.
In other words, we can say that the proposed model gives
better outcomes as the network become denser in size.
• As we know that the nature of the social network is dy-
namic, so one of the ubiquitous precincts of this article is
that the proposed approach is only suitable for the static
network. The main reason behind this boundary is the lack
of ground truth about the opinion leaders in the dynamic
social network datasets. In the future, we would take the
challenge to overcome this issue.
• Another limitation is that we have considered only a few
centrality measures that are used to evaluate the distance
parameter. As the new researches emerge day by day, we
have a higher opportunity to draw on other centrality mea-
sures in the future.
6. Conclusion and future scope
The social network has an excellent opportunity to identify the
consistent and authorized person who has the quality to persuade
the decision making power of other users in the network. It is a
very complicated task to identify such types of people in online
social networks due to various constraints such as fake user iden-
tity, user location, network structure, and many more. We can
also integrate user sentiment, tweets posted by the user, response
time, user’s location, recommendation and many other factors
embedded with the game theory-based approach to finding the
opinion leader more concisely [84–86]
In this paper, we have addressed a game theory-based ap-
proach to finding the coalition of opinion leaders based on a
maximum collective synergy of the group. We also recommended
the three types of trust to measure the degree of trust of a user.
Shapley value is also exploiting to find the average marginal
contribution of a user in a coalition utilized to produce synergy.
A user probably interacts with the others based on the degree
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158 17

Fig. 10. Comparative analysis of (a) Wiki-vote dataset (b) Bitcoin OTC trust weighted dataset, based on accuracy, precision, recall, and F1-score with other SNA
measures.

Fig. 11. Comparison of execution time for Wiki-vote dataset and Bitcoin OTC trust weighted dataset with other SNA measures.

Fig. 12. Impact of the total number of opinion leader on the diffusion rate in (a) Wiki-vote dataset (b) Bitcoin OTC trust weighted dataset.

of trust, and likelihood conditional probability depends on the user in every coalition by providing the four alternative solu-
other user’s action. We also wisely calculated the payoff of each tions. Therefore, we uniquely combined the power of the social
18 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
networks with the logic of game theory along with trust and
conditional probability to precisely identify the opinion leaders
in the social network. The proposed approach is also too much
expert as compared to other SNA measures and produced the
enhanced results with around 90% accuracy and 94% precision.
The computation time of the algorithm is also reduced by 27% as
compared to other SNA measures.
The applicability of the proposed method is very extensive as
nowadays, every organization wants to create wealth in the com-
mercial market. Each company desires that the sale of its products
would be high in the future. So opinion leaders have significant
exposure to achieve this goal. Not only in the global industry but
the other fields as education, healthcare, agriculture, shopping,
population controlling, and many more also. Recently, lots of
researches have been developed to illustrate the felicitousness of
opinion leaders solve many real-world problems [9,87–90].
The suggested approach also works as an intellectual expert
system that intelligently identifies the list of opinion leaders
using the game theory approach in the social network. An ex-
pert system is a system that can provide a solution to various
multifarious decision-making tasks using facts and rules [91]. The
proposed approach also followed the different ways of implemen-
tation and categorized the opinion leaders based on the synergy
of the alliance group. There are numerous real-world appliances
of opinion leaders, but we have discussed only a few of them in
brief. An opinion leader is very much efficient in the edification
domain to promote authenticate, reliable, and credible study
material for needy students. Another assistance of opinion leaders
is in the health sector to identify the experienced, knowledgeable,
honored, and skilled doctors and surgeons for the patients.
Further, most of the industries hire opinion leaders to pro-
mote their commodities inside the real world and also uses the
expertise of opinion leaders to analyze the feedback given by the
reviewers and other promoters. So, they are very much helpful
in increasing the wholesale price and gross growth rate of the
product. In a nutshell, the role of an opinion leader is genuinely
deserving and prominent for the diffusion of new products and
also can persuade the assessment by their convincing power and
strategy. In the future, we can also utilize the power of other
centrality measures, and game theory approaches to find the
list of opinion leaders. In the future, we may also recommend
the amalgamation of computational intelligence based techniques
with the social network dynamics to produce a large amount of
productive and creative opportunities in the different domains of
researches [19,92]. We might also apply the innovative computa-
tional intelligence techniques along with the evolutionary game
theory approach to detect the promising opinion leader in social
networks [93,94].
CRediT authorship contribution statement
Lokesh Jain: Conceptualization, Data curation, Formal analy-
sis, Funding acquisition, Investigation, Methodology, Project ad-
ministration, Resources, Software, Supervision, Validation, Visu-
alization, Writing - original draft, Writing - review & editing.
Rahul Katarya: Conceptualization, Data curation, Formal analy-
sis, Funding acquisition, Investigation, Methodology, Project ad-
ministration, Resources, Software, Supervision, Validation, Visu-
alization, Writing - original draft, Writing - review & editing.
Shelly Sachdeva: Conceptualization, Data curation, Formal anal-
ysis, Funding acquisition, Investigation, Methodology, Project ad-
ministration, Resources, Software, Supervision, Validation, Visu-
alization, Writing - original draft, Writing - review & editing.
Declaration of competing interest
The authors declare that they have no known competing finan-
cial interests or personal relationships that could have appeared
to influence the work reported in this paper.
References
[1] V. Buskens, Social networks and trust, Theory Decis. (2002) http://dx.doi.
org/10.1007/0-306-47645-2_1.
[2] K.M. Carley, Dynamic network analysis, Dyn. Soc. Netw. Model. Anal. Work.
Summ. Pap. (2003) http://dx.doi.org/10.1103/PhysRevB.81.041203.
[3] C. Kadushin, Introduction to social network theory, Networks (2002) http:
//dx.doi.org/10.1007/978-1-4614-2254-9.
[4] Q. Miao, S. Zhang, Y. Meng, H. Yu, Domain-sensitive opinion leader mining
from online review communities, in: Proc. 22nd Int. Conf. World Wide
Web - WWW ’13 Companion, 2013, pp. 187–188, http://dx.doi.org/10.
1145/2487788.2487882.
[5] P. Parau, C. Lemnaru, M. Dinsoreanu, R. Potolea, Chapter 10 - opinion
leader detection, in: Sentim. Anal. Soc. Networks, 2017, http://dx.doi.org/
10.1016/B978-0-12-804412-4.00010-3.
[6] K.K. Chan, S. Misra, Characteristics of the opinion leader: A new dimension,
J. Advert. (1990) http://dx.doi.org/10.1080/00913367.1990.10673192.
[7] M.C. Nisbet, J.E. Kotcher, A two-step flow of influence?: Opinion-leader
campaigns on climate change, Sci. Commun. 30 (2009) 328–354, http:
//dx.doi.org/10.1177/1075547008328797.
[8] S.M.H. Bamakan, I. Nurgaliev, Q. Qu, Opinion leader detection: A method-
ological review, Expert Syst. Appl. (2019) http://dx.doi.org/10.1016/j.eswa.
2018.07.069.
[9] H.C. Lin, P.F. Bruning, H. Swarna, Using online opinion leaders to promote
the hedonic and utilitarian value of products and services, Bus. Horiz.
(2018) http://dx.doi.org/10.1016/j.bushor.2018.01.010.
[10] F. Meng, J. Wei, Q. Zhu, Study on the impacts of opinion leader in online
consuming decision, in: Proc. - 2011 Int. Jt. Conf. Serv. Sci. IJCSS 2011,
2011, http://dx.doi.org/10.1109/IJCSS.2011.79.
[11] D. Lee, Game theory and neural basis of social decision making, Nat.
Neurosci. (2008) http://dx.doi.org/10.1038/nn2065.
[12] P.P. Shenoy, On coalition formation: a game-theoretical approach, Int. J.
Game Theory (1979) http://dx.doi.org/10.1007/BF01770064.
[13] D. Ray, A game-theoretic perspective on coalition formation, 2007, http:
//dx.doi.org/10.1093/acprof:oso/9780199207954.001.0001.
[14] Y. Narahari, R. Narayanam, Tutorial: Game theoretic models for social
network analysis, in: Proc. 20th Int. Conf. Companion World Wide Web,
WWW 2011, 2011, http://dx.doi.org/10.1145/1963192.1963316.
[15] M. del Pozo, C. Manuel, E. González-Arangüena, G. Owen, Centrality in
directed social networks. A game theoretic approach, Soc. Netw. (2011)
http://dx.doi.org/10.1016/j.socnet.2011.04.001.
[16] M.H. Manshaei, Q. Zhu, T. Alpcan, T. Basar, J.P. Hubaux, Game theory
meets network security and privacy, ACM Comput. Surv. (2013) http:
//dx.doi.org/10.1145/2480741.2480742.
[17] D. Gómez, E. González-Arangüena, C. Manuel, G. Owen, M. del Pozo,
J. Tejada, Centrality and power in social networks: A game theoretic
approach, Math. Soc. Sci. (2003) http://dx.doi.org/10.1016/S0165-4896(03)
00028-3.
[18] V. Mak, The emergence of opinion leaders in social networks, SSRN
Electron. J. (2008) http://dx.doi.org/10.2139/ssrn.1157285.
[19] Y. Zhao, G. Kou, Y. Peng, Y. Chen, Understanding influence power of opinion
leaders in e-commerce networks: An opinion dynamics theory perspective,
Inf. Sci. (Ny). 426 (2018) 131–147, http://dx.doi.org/10.1016/j.ins.2017.10.
031.
[20] L. Jain, R. Katarya, A systematic survey of opinion leader in online social
network, in: ICSNS 2018 - Proc. IEEE Int. Conf. Soft-Computing Netw.
Secur., 2018, pp. 1–5, http://dx.doi.org/10.1109/ICSNS.2018.8573639.
[21] A.K. Jagannatham, V. Kumar, Introduction to game theory, in: Decis. Sci.
Theory Pract., 2016, http://dx.doi.org/10.1201/9781315183176.
[22] D.M.W. Powers, Evaluation: from precision, recall and f-measure to roc,
informedness, markedness and correlation, J. Mach. Learn. Technol. (2011)
doi: 10.1.1.214.9232.
[23] F. Bodendorf, C. Kaiser, Detecting opinion leaders and trends in online
social networks, in: Proc. 2nd ACM Work. Soc. Web Search Min., 2009,
http://dx.doi.org/10.1109/ICDS.2010.29.
[24] F. Bodendorf, C. Kaiser, Detecting opinion leaders and trends in online
communities, in: 4th Int. Conf. Digit. Soc. ICDS 2010, Incl. CYBERLAWS
2010 1st Int. Conf. Tech. Leg. Asp. E-Society, 2010, pp. 124–129, http:
//dx.doi.org/10.1109/ICDS.2010.29.
[25] S. Al-Oufi, H.N. Kim, A. El Saddik, A group trust metric for identifying
people of trust in online social networks, Expert Syst. Appl. 39 (2012)
13173–13181, http://dx.doi.org/10.1016/j.eswa.2012.05.084.
 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
[26] Y. Cho, J. Hwang, D. Lee, J. Wang, D. Lee, Identification of effective opinion
leaders in the diffusion of technological innovation: A social network
approach, Technol. Forecast. Soc. Change (2012) http://dx.doi.org/10.1016/
j.techfore.2011.06.003.
[27] Y. Li, S. Ma, Y. Zhang, R. Huang, Kinshuk, An improved mix framework for
opinion leader identification in online learning communities, Knowl.-Based
Syst. 43 (2013) 43–51, http://dx.doi.org/10.1016/j.knosys.2013.01.005.
[28] M.F. Tsai, C.W. Tzeng, Z.L. Lin, A.L.P. Chen, Discovering leaders from social
network by action cascade, Soc. Netw. Anal. Min. 4 (2014) 1–10, http:
//dx.doi.org/10.1007/s13278-014-0165-9.
[29] J. Duan, J. Zeng, B. Luo, Identification of opinion leaders based on user
clustering and sentiment analysis, in: Proc. - 2014 IEEE/WIC/ACM Int. Jt.
Conf. Web Intell. Intell. Agent Technol. - Work. WI-IAT 2014, 2014, pp.
377–383, http://dx.doi.org/10.1109/WI-IAT.2014.59.
[30] A. Aleahmad, P. Karisani, M. Rahgozar, F. Oroumchian, Olfinder: Finding
opinion leaders in online social networks, J. Inf. Sci. 42 (2016) 656–674,
http://dx.doi.org/10.1177/0165551515605217.
[31] G. Sun, S. Bin, A new opinion leaders detecting algorithm in multi-
relationship online social networks, Multimedia Tools Appl. 77 (2018)
4295–4307, http://dx.doi.org/10.1007/s11042-017-4766-y.
[32] S.M. Aghdam, N. Jafari Navimipour, Opinion leaders selection in the social
networks based on trust relationships propagation, Karbala Int. J. Mod. Sci.
2 (2016) 88–97, http://dx.doi.org/10.1016/j.kijoms.2016.02.002.
[33] Y. Chen, L. Hui, C.I. Wu, H. Liu, S. Chen, Opinion leaders discovery in
dynamic social network, 2017.
[34] Y.C. Chen, A novel algorithm for mining opinion leaders in social networks,
World Wide Web (2019) http://dx.doi.org/10.1007/s11280-018-0586-x.
[35] F.K. Dewi, S.B. Yudhoatmojo, I. Budi, Identification of opinion leader on
rumor spreading in online social network twitter using edge weighting
and centrality measure weighting, in: 2017 12th Int. Conf. Digit. Inf. Manag.
ICDIM 2017, 2018, http://dx.doi.org/10.1109/ICDIM.2017.8244680.
[36] Y. Liu, Z. Gu, T.H. Ko, J. Liu, Identifying key opinion leaders in social media
via modality-consistent harmonized discriminant embedding, IEEE Trans.
Cybern. (2018) 1–12, http://dx.doi.org/10.1109/TCYB.2018.2871765.
[37] L. Yang, Y. Qiao, Z. Liu, J. Ma, X. Li, Identifying opinion leader nodes in
online social networks with a new closeness evaluation algorithm, Soft
Comput. 22 (2018) 453–464, http://dx.doi.org/10.1007/s00500-016-2335-
3.
[38] L. Yang, Y. Tian, J. Li, J. Ma, J. Zhang, Identifying opinion leaders in social
networks with topic limitation, Cluster Comput. 20 (2017) 2403–2413,
http://dx.doi.org/10.1007/s10586-017-0732-8.
[39] L. Jain, R. Katarya, Identification of opinion leader in online social network
using fuzzy trust system, in: Proc. 8th Int. Adv. Comput. Conf. IACC 2018,
2018, http://dx.doi.org/10.1109/IADCC.2018.8692095.
[40] Q. Liqing, D. Jinlong, L. Haiyan, W. Yan, Detecting opinion leaders in online
social networks using hybridrank algorithm, J. Intell. Fuzzy Syst. (2018)
http://dx.doi.org/10.3233/JIFS-169607.
[41] Z. Bu, H.J. Li, C. Zhang, J. Cao, A. Li, Y. Shi, Graph k-means based on leader
identification, dynamic game and opinion dynamics, IEEE Trans. Knowl.
Data Eng. (2019) http://dx.doi.org/10.1109/TKDE.2019.2903712.
[42] F. Riquelme, P. Gonzalez-Cantergiani, D. Hans, R. Villarroel, R. Munoz, Iden-
tifying opinion leaders on social networks through milestones definition,
IEEE Access (2019) http://dx.doi.org/10.1109/ACCESS.2019.2922155.
[43] S. Lei, X. Zhang, Y. Yan, Research on opinion leaders recognition based
on TOPSIS in open source design community, in: Proc. 2019 IEEE 23rd
Int. Conf. Comput. Support. Coop. Work Des. CSCWD 2019, 2019, http:
//dx.doi.org/10.1109/CSCWD.2019.8791850.
[44] L. Jain, R. Katarya, Discover opinion leader in online social network using
firefly algorithm, Expert Syst. Appl. (2019) http://dx.doi.org/10.1016/j.eswa.
2018.12.043.
[45] J. Yang, Y. Zhang, L. Liu, Identifying Opinion Leaders in Virtual Travel
Community Based on Social Network Analysis, in: Lect. Notes Comput.
Sci. (Including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics),
2019, http://dx.doi.org/10.1007/978-3-030-22338-0_23.
[46] L. Jain, R. Katarya, S. Sachdeva, Opinion leader detection using whale
optimization algorithm in online social network, Expert Syst. Appl. (2019)
113016, http://dx.doi.org/10.1016/j.eswa.2019.113016.
[47] A.U. Rehman, A. Jiang, A. Rehman, A. Paul, S. din, M.T. Sadiq, Identification
and role of opinion leaders in information diffusion for online discussion
network, J. Ambient Intell. Humaniz. Comput. (2020) http://dx.doi.org/10.
1007/s12652-019-01623-5.
[48] S. Walter, M. Brüggemann, Opportunity makes opinion leaders: analyzing
the role of first-hand information in opinion leadership in social media
networks, Inf. Commun. Soc. (2020) http://dx.doi.org/10.1080/1369118X.
2018.1500622.
[49] Z. Wang, H. Liu, W. Liu, S. Wang, Understanding the power of opinion
leaders’ influence on the diffusion process of popular mobile games: Travel
Frog on Sina Weibo, Comput. Hum. Behav. (2020) http://dx.doi.org/10.
1016/j.chb.2020.106354.
19
[50] J. Cao, Z. Bu, Y. Wang, H. Yang, J. Jiang, H.J. Li, Detecting prosumer-
community groups in smart grids from the multiagent perspective, IEEE
Trans. Syst. Man Cybern. Syst. (2019) http://dx.doi.org/10.1109/TSMC.2019.
2899366.
[51] I.H. Witten, E. Frank, M.A. Hall, C.J. Pal, Data mining: Practical machine
learning tools and techniques, 2016, http://dx.doi.org/10.1016/c2009-0-
19715-5.
[52] J. Han, M. Kamber, J. Pei, Data mining: Concepts and techniques, 2012,
http://dx.doi.org/10.1016/C2009-0-61819-5.
[53] M. Kwan, D. Ramachandran, Trust and online reputation systems, 2009,
pp. 287–311, http://dx.doi.org/10.1007/978-1-84800-356-9_11.
[54] L. Liu, W. Shi, Trust and reputation management, IEEE Internet Comput.
(2010) http://dx.doi.org/10.1109/MIC.2010.124.
[55] Y. Ruan, A. Durresi, A survey of trust management systems for online social
communities - trust modeling, trust inference and attacks, Knowl.-Based
Syst. 106 (2016) 150–163, http://dx.doi.org/10.1016/j.knosys.2016.05.042.
[56] M. Steketee, A. Miyaoka, M. Spiegelman, Social network analysis, in: Int.
Encycl. Soc. Behav. Sci., second ed., 2015, http://dx.doi.org/10.1016/B978-
0-08-097086-8.10563-X.
[57] T. Schank, D. Wagner, Approximating clustering coefficient and transitivity,
J. Graph Algorithms Appl. (2005) http://dx.doi.org/10.7155/jgaa.00108.
[58] M. McPherson, L. Smith-Lovin, J.M. Cook, Birds of a feather: Homophily in
social networks, Annu. Rev. Sociol. 27 (2001) 415–444, http://dx.doi.org/
10.1146/annurev.soc.27.1.415.
[59] S. Zhang, Y. Zhang, Introduction to game theory, Chin. Sci. Bull. (2003)
http://dx.doi.org/10.1360/99ws0111.
[60] R. Leonard, Game theory, in: Handb. Hist. Econ. Anal., 2016, http://dx.doi.
org/10.4018/jaci.2011070106.
[61] W.P. Fox, R. Burks, Game theory, in: Int. Ser. Oper. Res. Manag. Sci, 2019,
http://dx.doi.org/10.1007/978-3-030-20569-0_6.
[62] K. Leyton-Brown, Y. Shoham, Essentials of game theory: A concise mul-
tidisciplinary introduction, Synth. Lect. Artif. Intell. Mach. Learn. (2008)
http://dx.doi.org/10.2200/s00108ed1v01y200802aim003.
[63] C.F. Camerer, Behavioral game theory: Experiments in strategic interaction,
2003, http://dx.doi.org/10.1016/j.socec.2003.10.009.
[64] J. von Neumann, O. Morgenstern, Theory of games and economic behavior,
2007, http://dx.doi.org/10.2307/3610940.
[65] W. Thomson, A.E. Roth, The shapley value: Essays in honor of Lloyd S.
Shapley, Economica (1991) http://dx.doi.org/10.2307/2554979.
[66] L.S. Shapley, 17. A value for n-person games, in: Contrib. To Theory Games
(AM-28), II, 2016, http://dx.doi.org/10.1515/9781400881970-018.
[67] T. Ui, A shapley value representation of potential games, Games Econ.
Behav. (2000) http://dx.doi.org/10.1006/game.1999.0729.
[68] R. Narayanam, Y. Narahari, A shapley value-based approach to discover
influential nodes in social networks, IEEE Trans. Autom. Sci. Eng. (2011)
http://dx.doi.org/10.1109/TASE.2010.2052042.
[69] T.P. Michalak, K.V. Aadithya, P.L. Szczepanski, B. Ravindran, N.R. Jennings,
Efficient computation of the shapley value for game-theoretic network
centrality, J. Artificial Intelligence Res. (2013) http://dx.doi.org/10.1613/jair.
3806.
[70] E. Hernandez, J.M. Shaver, Network synergy, Adm. Sci. Q. (2019) http:
//dx.doi.org/10.1177/0001839218761369.
[71] V. Buskens, The social structure of trust, Soc. Netw. (1998) http://dx.doi.
org/10.1016/S0378-8733(98)00005-7.
[72] S. Sanadhya, S. Singh, Trust Calculation with ant colony optimization in
online social networks, Procedia Comput. Sci. 54 (2015) 186–195, http:
//dx.doi.org/10.1016/j.procs.2015.06.021.
[73] B.A. Brooks, The strength of weak ties, Nurse Lead. (2019) http://dx.doi.
org/10.1016/j.mnl.2018.12.011.
[74] M.S. Granovetter, The strength of weak ties, Am. J. Sociol. (1973) http:
//dx.doi.org/10.1086/225469.
[75] R.S. Burt, Structural holes and good ideas, Am. J. Sociol. (2004) http:
//dx.doi.org/10.1086/421787.
[76] S. Goyal, F. Vega-Redondo, Structural holes in social networks, J. Econ.
Theory. (2007) http://dx.doi.org/10.1016/j.jet.2007.01.006.
[77] M. Bastian, S. Heymann, M. Jacomy, Gephi: An open source software for
exploring and manipulating networks, in: Third Int. AAAI Conf. Weblogs
Soc. Media, 2009, http://dx.doi.org/10.1136/qshc.2004.010033.
[78] M.F. Sanner, Python: A programming language for software integration and
development, J. Mol. Graph. Model. (1999).
[79] J. Leskovec, D. Huttenlocher, J. Kleinberg, Predicting positive and negative
links in online social networks, in: Proc. 19th Int. Conf. World Wide Web,
WWW ’10, 2010, http://dx.doi.org/10.1145/1772690.1772756.
[80] S. Kumar, F. Spezzano, V.S. Subrahmanian, C. Faloutsos, Edge weight
prediction in weighted signed networks, in: Proc. - IEEE Int. Conf. Data
Mining, ICDM, 2017, http://dx.doi.org/10.1109/ICDM.2016.175.
[81] T. Fawcett, An introduction to ROC analysis, Pattern Recognit. Lett. (2006)
http://dx.doi.org/10.1016/j.patrec.2005.10.010.
[82] P.S. Van Eck, W. Jager, P.S.H. Leeflang, Opinion leaders’ role in innovation
diffusion: A simulation study, J. Prod. Innov. Manag. (2011) http://dx.doi.
org/10.1111/j.1540-5885.2011.00791.x.
20 L. Jain, R. Katarya and S. Sachdeva / Knowledge-Based Systems 203 (2020) 106158
[83] Y. Cho, J. Hwang, D. Lee, Identification of effective opinion leaders in the
diffusion of technological innovation: A social network approach, Technol.
Forecast. Soc. Change (2012) http://dx.doi.org/10.1016/j.techfore.2011.06.
003.
[84] B. Liu, Sentiment analysis and opinion mining, Synth. Lect. Hum. Lang.
Technol. (2012) http://dx.doi.org/10.2200/S00416ED1V01Y201204HLT016.
[85] Encyclopedia of social network analysis and mining, 2014, http://dx.doi.
org/10.1007/978-1-4614-6170-8.
[86] K. Xu, S. Zhang, H. Chen, H.T. Li, Measurement and analysis of online
social networks, Jisuanji Xuebao/Chin. J. Comput. (2014) http://dx.doi.org/
10.3724/SP.J.1016.2014.00165.
[87] L. Jain, R. Katarya, S. Sachdeva, Role of opinion leader for the diffusion of
products using epidemic model in online social network, in: 2019 Twelfth
Int. Conf. Contemp. Comput, 2019, pp. 1–6, http://dx.doi.org/10.1109/IC3.
2019.8844911.
[88] T.W. Valente, P. Pumpuang, Identifying opinion leaders to promote
behavior change, Health Educ. Behav. (2007) http://dx.doi.org/10.1177/
1090198106297855.
[89] L. Locock, S. Dopson, D. Chambers, J. Gabbay, Understanding the role of
opinion leaders in improving clinical effectiveness, Soc. Sci. Med. (2001)
http://dx.doi.org/10.1016/S0277-9536(00)00387-7.
[90] N. Keys, D.C. Thomsen, T.F. Smith, Opinion leaders and complex sustain-
ability issues, Manag. Environ. Qual. Int. J. (2010) http://dx.doi.org/10.1108/
14777831011025535.
[91] Fundamentals of Expert System, in: Springer Ser. Mater. Sci, 2015, http:
//dx.doi.org/10.1007/978-3-662-44497-9_3.
[92] B. Skyrms, R. Pemantle, A dynamic model of social network formation,
Proc. Natl. Acad. Sci. USA 97 (2000) 9340–9346, http://dx.doi.org/10.1073/
pnas.97.16.9340.
[93] R. Cressman, J. Apaloo, Evolutionary game theory, in: Handb. Dyn. Game
Theory, 2018, http://dx.doi.org/10.1007/978-3-319-44374-4_6.
[94] Computational intelligence: a logical approach, Choice Rev. Online (1998)
http://dx.doi.org/10.5860/choice.35-5701.