1. Introduction
With the rise of communication technology, people are utilizing platforms such as content sharing sites, blogs, social networking, and wikis to create, modify, share, and discuss Internet content. Social media provides flexible platforms that play key roles in energizing collective action in movements [
1]. This represents the social media phenomenon, which can significantly impact society and industry, e.g., firms’ reputations, sales, and even survival [
2]. Within the discussions on social media, certain individuals influence others and thus emerge as opinion leaders. Opinion leaders have great impacts and influence on social media. Organizations can take advantage of these predispositions through marketing research and public relations, nurturing opinion leaders or advocates, placing and creating advertisements, developing new products and lowering the cost-to-serve [
3]. On the Internet, the power of these leaders is increasing larger and sequentially influencing entire societies through calls to protest, promotion of policy and decision-making, which was defined as the fifth right in the “Towards a Civil Society” seminar [
4].
The world must confront the energy crisis and air pollution. Discussions about energy issues are increasing. These discussions range from nuclear energy, thermal power, hydropower, and other forms of green energy and low carbon technology, including wind, solar, tidal, and biomass geothermal energy issues. In Taiwan, whenever an energy crisis occurs, energy charges increase. Anti-nuclear positions and other energy issues are discussed broadly. Therefore, the Taiwan government tries to understand people’s needs and questions.
On the Internet, the roles of opinion leaders and followers in the formation of these issues cannot be neglected. According to the “theory of two-step flow” [
5] and Rosen’s definition of opinion leaders’ characteristics, “social media initially pass the information to opinion leaders, then opinion leaders spread the information to followers and influence their attitudes” [
6]. Thus, when followers follow opinion leaders, the formers’ judgments and attitudes will be influenced and changed by opinion leaders. This study defines opinion leaders as people or social media with high social status who are able to influence followers. This study defines followers as the users who follow certain issues, publish related discussions and add their own ideas. They spread, repost or blindly follow the behaviors of opinion leaders.
Most previous research of opinion leaders focuses on the commercial domain rather than on nonprofit-related policies such as energy policy [
7]. In the promotion of many public policies through online postings, it is difficult to clearly identify opinion leaders and followers, which greatly reduces the effectiveness of communication. Based on the community attributes of opinion leaders and whether they can successfully resonate, this study aims at providing a method to try to identify who are opinion leaders or who are likely to become opinion leaders in social media, and who are followers. Relational matrix analysis is used to represent the relationship between opinion leaders and followers in social media and to identify the collection of opinion leaders and potential opinion leaders.
Furthermore, previous studies [
8] have used quantitative methods of analysis. One example is the SuperedgeRank algorithm. However, this algorithm not only has difficulty identifying potential opinion leaders effectively but also neglects how opinion leaders influence followers and how relationships between opinion leaders and followers are characterized. It also ignores the increasingly important role played by intelligent systems such as algorithms.
Although the literature on green energy is rapidly increasing, many studies suggest that this problem needs to be dealt with by considering a broader perspective [
9]. This study not only examines the issue from the perspective of intelligent systems such as algorithms but also identifies the roles of opinion leaders and followers on social media in relation to the introduction of green energy and carbon reduction technology, with the aim of developing public policy to promote green energy and low carbon emissions. This study is novel not only because it takes quantitative factors and tradition clustering approaches into account but because it also analyzes posts, poster characteristics and their interactive relationships on social media. This study reviews relevant literature in
Section 2. In
Section 3, we propose a method to identify opinion leaders and their followers based on their interactions on social networks. The interaction patterns are also identified. An energy case is studied in
Section 4 to validate the proposed solution approach and enhance communication effectiveness between government policymakers and people’s desires. The discussion is summarized in
Section 5, and
Section 6 concludes this study.
This research contributes to finding a solution to easily identify the characteristics of opinion leaders and followers in the case of online posts related to green energy and low-carbon policies. Once certain public policies need to be effectively disseminated, they can be widely used. Using the same model and the solution approach, the results of this study can be extended from the green energy low carbon issue to other social issues. Furthermore, this study provides a new perspective to deal with the effective identification of opinion leaders and followers, at the same time, promotes the “theory of two-step flow” to add another research perspective in the academic field.
3. Methodology
3.1. Modeling Interaction between Opinion Leaders and Followers
Between users, matrix
M is a relational matrix.
We set the row index as i and the column index as j in the matrix. The n users in the set are represented by C = {1 … n}. In the matrix T, the elements are composed of counts of responses and being responded to. Additionally, between users and social community support level, the elements in the matrix I are the influence factors
Matrix T shows us the counts of both responses and being responded to between n users. T_ij refers to counts of responses of useri_i to userj_j where i ≠ j. When T_ij refers to counts of total posts of user_i where i = j. i refers to the index of the users who responds to other user’s opinions, and j refers to the index of the users whose opinions are responded to.
Matrix I indicates the power of users to influence and be influenced.
I_ij refers to the influence of
user_i on
user_j where
i ≠ j,
i refers to the index of a user who influences other people, and
j refers to the index of a user who is influenced by others. The influence power can be classified into three different patterns, including job position, professional knowledge and social community support level [
39]. Each criterion has three levels, no influence (NI), general influence (GI), and high influence (HI) (
Table 1). For a higher job title with more professional knowledge and a high social community support level, we define influence power as high influence (HI). For a general job title with popular professional knowledge and neutral social community support level, the overall influence power is general (GI). If the user has no job and inaccurate professional knowledge or social community support, we define that the influence level as having no influence (NI). However, some users have privacy settings or discuss issues anonymously, so our study could not gather their background information. Fortunately, the proposed approach can serve as a flexible model with the missing part left blank.
In matrix I, I_ij refers to a user’s social community support level where i = j. In this study, social community support is divided into three levels. These three levels are relatively well-known, and well-followed social media that receives good attention are which are classified as having high social community support (H). Less-followed, less-known social media with low attention is classified as low (L). However, we were unable to gather users’ social community support levels because of some users’ anonymous discussions or privacy settings. Our study defines the social community support level of these users as missing (O).
A non–follower, a user with a negative speech count, is defined in this study. Meanwhile, when contents are responded to negatively, the user is listed as a non-opinion leader. The relationship between opinion leaders and followers refers to a mutual relationship between users. However, one user may not respond to or express ideas to others’ speech content on social media. One opinion leader may not be an opinion leader of all users. Therefore, if there is no relationship between users, we cannot distinguish whether they are opinion leaders or followers. In this study, the groups of opinion leaders and followers are be judged base on interaction. The users with mutual influence are classified as one group to analyze whether there is an opinion leader and a follower in the group. If there is no mutual influence, no group is formed.
3.2. Opinion Leaders and Followers’ Social Patterns
Six axioms are proposed to classify opinion leaders, influencers, followers and interaction patterns. Notations are presented in
Table 2.
The axioms formulated by the systematic rolling analysis of big data over the years. Experts determine the criteria and select thresholds for them, for example, the positive or negative degree of the counts of responses to and responded to. The following axioms are proposed to recognize opinion leaders, influencers, and followers.
Axiom1—Opinion leader.
If and and
then the
is an opinion leader.
To be an opinion leader, ’s () response total must be higher than the average total response (), and must have more than twice the number of responses in the group (). Moreover, speech and the count in matrix T() less than 1 cannot be more than half of the total number of members’ responses in group g.
Axiom 2—Influencer.
If , then is an influencer. If the responses () are more than the average number of total responses in the group and lower than twice the average number of total responses in the group. Moreover, speech which is NI and the count in matrix T() which is less than 1 cannot more than the number of the member which is in the group. Influencers also have the power to influence others and have the potential to become opinion leaders. Hence, this study also offers a way to find influencers.
Axiom 3—Followers.
If , then the user is a follower.
A follower needs to support or agree with someone, so the must have a positive count in matrix T. In other words, the user in is the follower’s leader.
Three social community patterns are classified: In the criterion pattern, the opinion leaders broadly influencing many followers usually obtain high social community support and posts professionally. The criterion social community, the most common pattern. In the argument pattern, the pattern’s emergence is caused by the discussion space provided for users of social community platforms. Users can give specific advice to influence each other or influence other users. The bandwagon pattern arises when followers follow closely due to opinion leaders’ personal charisma. In this pattern, followers usually do not care whether the content posted by opinion leaders is correct.
The following axioms are used to define three patterns.
Axiom 4—Criterion pattern.
According to group_k with opinion leaders in set_ol, (1) if the number of posts influencing followers who post to set_ol in group_k is more than half of the number (≥1/2 × SI_f), and (2) if the number of posts influencing opinions directed to followers in group_k is more than half of the number (SI_ol {HI,GI} ≥ 1/2 × SI_ol), then group_k is a criterion pattern.
According to the criterion pattern, an enterprise can promote its products effectively, and the government can sharp public opinion in favor of a particular policy by utilizing the function of opinion leaders. Furthermore, finding opinion leaders and tracking them over the long term can prevent an explosion of potential issues. In the criterion pattern, the opinion leader is very professional. If a government or enterprise wants to negotiate or cooperate with them, the contractor must also be professional.
Axiom 5—Argument pattern.
After grouping with the CI Algorithm, we find different groups. According to group_k with opinion leaders in set_ol, (1) if the number of posts influencing {HI,GI} followers who post to set_ol in group_k is more than half of the number (if SI_f {GI,HI} ≥ 1/2 × SI_f) and (2) if the number of posts influencing {HI, GI} opinions directed to followers in group_k is more than half of the number (SI_ol {HI,GI} ≥ 1/2 × SI_ol), then we recognize that group_k can be characterized as an argument pattern.
On the basis of the argument pattern, the character of the interaction between opinion leaders and followers is not significant; consequently, the cost of marketing is high and may even have little impact on promotion. Moreover, in the argument pattern, the viewpoints are diverse. Thus, it is desirable to provide a platform and sufficient information and domain knowledge for uses to engage in dialog with each other.
Axiom 6—Bandwagon pattern.
After grouping with the CI Algorithm, then we have different groups. According to group_k with opinion leaders in set_ol, (1) if the number of posts influencing {NI} followers who post to set_ol in group_k is more than half of the number (if SI_f {NI} ≥ 1/2 × SI_f) and (2) if the number of posts influencing {NI} opinions directed at followers in group_k is more than half of the number (SI_ol {NI} >1/2 × SI_ol), then we recognize that group_k can be characterized as a bandwagon pattern.
According to the bandwagon pattern, opinion leaders and followers are not professional in most cases. Enterprises and governments can utilize social media to promote their products and public policy effectively by enhancing the roles of these opinion leaders and followers. If the government or an enterprise wants to negotiate or cooperate with them, the contactors need not be professional but must be a decision-maker who can promise to provide resources.
3.3. Problem with Identification of Opinion Leaders and Followers
The problem of identifying opinion leaders and followers is formulated as follows:
Decompose a user-user interaction matrix into mutually separable submatrices (modules) with (1) the minimum number of non-empty high-value entries outside the block-diagonal matrix T, and (2) the maximum number of strongly desired entries (HI) and the minimum number of strongly undesired entries (NI) included in the submatrices of the block diagonal matrix I.
Subject to the following constraints:
Constraint C1: Empty groups of users are allowed, and
Constraint C2: The number of users in a group cannot exceed the upper bound Nu.
Constraint C3: Satisfy the following assumptions:
- (1)
Continuous posts are defined as one post.
- (2)
Users who respond negatively to posts cannot be regarded as followers.
- (3)
The content of the post and the level of social community support determine the influence of the user’s post.
- (4)
Expert posts are prioritized as reasonable posts.
- (5)
If users have a low influence on each other, judge it as “NI”.
- (6)
If users have a great influence on each other, judge it as “HI”.
In matrix T, we count input post, responses and being responded to. In the matrix I, (1) input the highest influence power HI. Moreover, (2) input the lowest influence power NI. Combined with observation results, identify the opinion leader set and follower set.
3.4. Identification of Opinion Leader and Follower
In this study, the relationship between opinion leaders and followers is the mutual relationship between users. When a user satisfies the characteristics of opinion leaders, our study defines the user as an opinion leader. Followers will change their own ideas and attitudes according to opinion leaders’ characteristics, including social status, accuracy of post contents and social community support level. The algorithm is described as follows:
Step I. Collect data from social media, such as users, posts, and response information.
Step II. Compute the counts of total posts of in matrix [Tii]
Step III. For all users, put the responses which gives to in matrix [Tij] until there are no responses from to other users.
Step IV. If and in matrix T are NULL then remove the meaningless by deleting and in matrix T and I until there are no meaningless
Step V. According to the data of social media, matrix T and social community support level, the social community support level marked with at [Iii].
Step VI. For all users i and j, according to expert judgment, assign the influence power level [Iij] of .
Step VII. If ] > [], exchange the columns of . until there is no [] <] in matrix T.
Step VIII. The CI algorithm is applied to group users.
Step IX. If [Iij] is not NULL, then check whether the and is in the same group or not; if they are not in the same group, then put them in the same group matrix until all users in the matrix I have been checked.
Step X. In each group, sum up all positive responses to and compute the average .
Step XI. Count the responses of by and posts of by . Additionally, count the influence of each , If satisfies Axiom 1, then is identified as an opinion leader. If satisfies Axiom 2, then is identified as an influencer. If satisfies Axiom 3, then is identified as a follower. Continue until all users in matrix T have been checked.
Step XII. Check each group matrix. Count all opinion leaders’ [] and followers’ []. Recognize the pattern based on Axioms 4–6.
Step XIII. When there are positive responses or influence between users, we classify these users in the same group.
4. Case Study
The ABC network platform is taken as an example to describe the application of our research in practice. The ABC network platform is a discussion platform created by the government to promote community communication. This platform was created as part of a public policy proposal to improve policy communication and make policy public.
This case study is taken from the National Energy Conference organized by the Energy Bureau of Taiwan’s Ministry of Economic Affairs. However, there are still many disagreements when it comes to choosing opinions due to value divergence. To discuss and clarify issues with the public, the proposition, “Where does future electric power come from?” is open on the policy consultant forum (People Talk), with three sub-issue forums including, ‘environment low carbon sustainable development’, ‘stable supply and open source’ and ‘reduce expenditures effectively’. In particular, ‘stable supply and open source’ is the focus of this case study.
The proposed solution approach is applied in this case.
Step I. Collect materials: Judging by the forum (posts, fan pages) on social media about green energy and low carbon, we collected materials, including text and response information. This study collected materials from users’ discussion contents related to the “stable supply” issue on the ABC network platform between May 2019 and the end of 2019. The data collection is implemented with the Python-Jieba crawler program, which is particularly suitable for Chinese text analysis automatically. The collected materials are listed below: Post users: 36; total posts: 205 (total posts have been deducted from the number of administrator responses and consecutive posts); effective responses: 61; effective count of being responded to 47.
Step II. Input matrix elements: Input elements in matrix T. The elements include general posts, counts of responses and being responded to as judged by experts.
Step III. Remove meaningless users: Remove users whose posts are never responded to.
Step IV. Tag the user’s category: Input social community support
Step V. Influence analysis: We analyze a user’s influence by comparing the levels of influence power according to three characteristics and input the influence into the matrix I.
Step VI. If the count of users’ posts, responses, and influence power, reaches a certain level of relevance, then move forward. If the users have greater counts of responses or respondents, list them in front. (
Figure 1).
Step VII. According to matrix I and T based on Equation (1), group the users by their relationships. Group A {1, 2, 3, 4, 5, 6, 18, 20, 21, 24}; Group B {2, 16, 24, 26, 27, 34, 36}; Group C {17, 18, 19}; Group D {19, 20, 21, 36}; Group E {2, 27, 28, 29, 30, 36}; Group F {7, 20, 26}; Group G {2, 9, 27, 31, 34}; Group H {7, 27}; Group I {7, 8, 15}; Group J {21, 22, 23}; Group K {25, 26}; Group L {34, 35}.
Step VIII. Identification: Determine the interaction between users, followers, and opinion leaders, according to the definition of each group of opinion leaders.
In Group A, opinion leader 1 is recognized by Axiom 1, and the influencers are Users 3 and 5. This group is identified as an argument group based on Axiom 5.
In Group B, User 16′s post contents are usually meaningless. According to Axiom 1, User 16 is not an opinion leader. Moreover, the group does not belong to any pattern.
Group C is an argument pattern. Users 18 and 19 are influencers. In this pattern, no opinion leaders and followers are identified. The influencers influence each other without focusing on any particular key person.
In group D, the response of User 19 has a great influence on Users 20 and 21. However, the influence of the posts responded to by Users 20 and 21 is not great. There is a discussion relationship between Users 19 and 20, so it is a social pattern.
In Group E, the post contents of User 36 are valuable. However, other users’ responses are not good. Thus, User 36 is an opposing opinion leader.
In Group F, according to Axiom 6, User 26 is an influencer, and the group represents an argument pattern. User 20 is the follower of User 26.
In Group G, Users 27 and 34 are the influencers, and it is a bandwagon pattern. Moreover, User 2 is the follower.
In Group H, Users 27 and 7 have a discussion relationship, and both of them are influencers.
In Group I, the social community support level of the group is high. They should be opinion leaders, in theory. However, in this case, study, since they do not play the role of opinion leaders, they cannot be recognized as opinion leaders. User 7 is an opinion leader, and this is an argument pattern.
In Group J, there are no opinion leaders or followers. User 21 is an influencer.
In Group K, User 26 is opposed to the opinions of User 25. There are no opinion leaders or followers in this group.
Group L: In this group, there are no opinion leaders or followers.
According to the summary of group analysis, Users 1 and 36 are obviously opinion leaders. The five groups A, C, D, F, and H can all be characterized as argument patterns, which shows that in the forum, most post contents influence other users through discussion.
The case was also analyzed with a traditional network approach, i.e., the Ward method, named after its creator, focuses on the allocation of profiles to groups equally. Ward [
50] pointed out that grouping in this manner makes it easier to consider and understand relations in large collections. The principle of this method is to minimize heterogeneity, and the important goal is to find the greatest similarity. The comparison between the proposed approach and Ward’s approach is shown in
Table 3. The results show that Users 1, 7, 36 are identified as opinion leaders. However, User 19 has not been identified through the traditional method due to the threshold.
The identification of opinion leaders by Ward’s method only identifies opinion leaders who participate in the whole conversation. However, in the proposed approach, this study uses two identification methods: the whole conversation and group conversation. The latter can clarify which user is the group’s opinion leader. In addition, the proposed approach can discover different patterns. Although the traditional network approach of Ward’s is considered to be the best one among the hierarchical clustering methods [
51,
52,
53], it cannot identify these patterns.
Through the perception of social community patterns among users, this study successfully distinguished opinion leader and defined social patterns in the complex social communities, which contains highly controversial users and many of them are anonymous, where few persons are involved in the discussion and users’ support level could not be obtained because users disagree with each other.
After identifying opinion leaders and followers, in a criterion pattern, opinion leaders have a higher degree of professionalism than followers. In that case, if green energy and low-carbon related policies are to be disseminated through opinion leaders, it is necessary to send personnel with a certain degree of professionalism. After contacting and negotiating with them, you must first obtain the approval of the opinion leaders before you can persuade them to influence followers through their platforms or social media. It is expected that they will achieve rapid dissemination, higher dissemination effect, and avoid costly but ineffective dissemination. In addition, a follower may also become another opinion leader, generating multiple diffusion of innovations.
Second, in the argument pattern, due to the comparably equal status between opinion leaders and followers, issues are quite diverse, and it is not easy to focus on specific issues. If opinion leaders wanted to disseminate relevant policies on energy conservation and low carbon to influence followers, the dissemination effect would be poor. Therefore, in order to make the issue of green energy and low-carbon attract more attention, opinion leaders can package the issue into lifestyles and features, thereby achieving a higher diffusion effect (Diffusion of innovations) on followers.
In addition, in the bandwagon pattern, because opinion leaders and followers are less professional, they are more vulnerable to each other. In order to disseminate green energy and low-carbon policies, policies can be packaged as simple, interesting or lifestyle issues, while social media or platforms are often used by opinion leaders or followers to achieve better diffusion of innovations.
5. Discussion
In this study, three interactive patterns and their characteristics are identified, which can help how to find opinion leaders more effectively and grasp the characteristics of opinion leaders and followers when want to spread (Diffusion of innovation) new policies or marketing new products. Opinion leaders and followers both have different levels of knowledge, social community support, and influence power. Therefore, this study summarizes the interactions on social media into three patterns, and the characteristics of three patterns have also been explored. Furthermore, based on the characteristics of users in these patterns, it can be used to provide opinion leaders with specific and clear topics/issues to influence their followers, thereby obtaining effective dissemination or commercial marketing purposes in the green energy domain.
In addition, the results of this study can also be applied to the political dissemination of democratic elections or the shaping of the opinion climate, which can more efficiently lead the electoral issues and win elections. In other words, the issues or political opinions that candidates are trying to market can be differentiated based on different communication modes so that the information can be segmented, and the impact of effective agenda-setting goals can be achieved. This study not only has the possibility of expanding and deeper research, but it is also the relative value of this research.
Most previous studies used different algorithms or improved algorithms to identify opinion leaders [
28,
29,
30,
31,
33,
37]. In other words, most of the above-mentioned studies only used various algorithms to identify opinion leaders or followers, and consequently, apply them to political communication and commercial marketing related fields. There has recently been an exploration of the psychological motivation of actively acting as opinion leaders to understand which users are active communicators or passive recipients of social issues. However, the related study on the interaction patterns between the opinion leaders and followers and their characteristics have not been explored.
In addition, the opinion dynamics of current popular research are interested. The classic model of opinion dynamics is derived from the research of DeGroot’s and Friedkin–Johnsen’s models of opinion dynamics, which aims at the integration and consistency of opinions in social networks, carried out very enlightening modeling and exploration [
54]. DeGroot’s model describes the process of reaching consensus in social networks, while Friedkin–Johnsen’s model further introduces the degree of “stubborn individual” to explain the phenomenon of inconsistent opinions in social networks. The models clearly depict the dynamic process of opinion integration and consistency, as well as the obstacles caused by “stubborn individual” factors to the process of opinion integration [
55,
56]. However, the two models are very instructive to explore how individuals (or Internet users) should be controlled if they are affected by certain characteristics or stubbornness in the process of opinion integration. Our study more specifically explores how opinion leaders and followers can find out the characteristics of users and the patterns of interaction between them in the process of consensus, which can be applied to precision marketing in actual operation, and even give opinion leaders with different characteristics use differentiated topic content to increase the influence of consensus. Since the research of DeGroot’s and Friedkin–Johnsen’s models of opinion dynamics are in conceptual level only, some difficulties in practical application are challenged [
57].
This study extends the idea of [
39] to the domain of green energy and low carbons, where roughly qualitative characteristics of opinion leaders and matrix of interaction between users are considered. To make this study more solid and applicable, the theories of “two-step flow”, “bandwagon effect”, “agenda-setting” and “innovation diffusion theory” from the theoretical perspective of communication and the axioms are used to validate the results. In addition, this study does not focus on a single discipline only but a cross-disciplinary study of the fields of green energy and low carbons, intelligent systems, and communication to provide numerous management implications discussed in this section. The core novelty and contribution is shown in that the solid theoretical part makes this study applicable to other social media and industry sectors.