Effects of Policy Communication Changes on Social Media: Before and After Policy Adjustment

Abstract

The structure of a policy communication network shows the effect of policy communication on social media. Policies need to be dynamically adjusted during the implementation process, which may affect the policy’s interaction on social media. Based on the Policy Network Theory, this study explores the effects of policy communication changes on social media before and after the adjustment of China’s Mass Entrepreneurship and Innovation (MEI) Policy using Exponential Random Graph Models (ERGMs) analysis and community analysis. The study reveals that after the policy adjustment, the communication network structure indicated a significant increase in triangular configurations, yet the formation of edges remained constrained. Meanwhile, cross-community connections in the communication network decreased, with communities exhibiting localized contraction, and emotional polarization becoming more pronounced. These phenomena occurred because policy adjustments have boosted interaction levels through new incentive mechanisms, whereas the content and delivery methods of policy communication remain insufficiently engaging, which constrains relationship-building. Additionally, the policy’s evolution from a mobilization–participation model to a vertical governance paradigm has systematically reconfigured inter-community interaction patterns, resulting in structural transformations in cross-group information flows. To enhance the dissemination of policies on social media, it is recommended to intervene in the policy communication network structure through role embedding, shift from a reactive public sentiment management paradigm to proactive emotional governance, and strengthen policy communication strategies that emphasize emotional resonance. These measures can improve the effectiveness of policy communication and help address the challenges posed by emotional polarization and network fragmentation.

1. Introduction

Policy implementation is a dynamic process [1]; flexibility and agility are created when implementing new or revised policies through policy adjustment or optimization [2]. Government and other institutions take on the role of explaining substantial differences that emerge during policy adjustments to ensure effective policy implementation [3]. Policy audience response is also crucial in policy implementation and adjustment [4]. Social media has become extremely popular [5] and has brought great difficulties, challenges, and opportunities for policy communication, making it possible for audiences to participate in policy formulation and implementation [6] by freely expressing their views through interactions [7]. Specifically, sentiment analysis on social media regarding specific events can provide better insights for planning and decision-making [8]. Social media has increasingly emerged as a pivotal platform for fostering government transparency, facilitating dialogue, and enhancing engagement with the public [9]. The informal evaluation and feedback of the policy audience on social media after policy adjustment provides decision makers with a reference and basis for assessment. Therefore, it is important to conduct a comparative study of the effect of interaction on social media before and after policy adjustments.

Policies gradually spread on social media through interaction, and the interactive network formed by users in policy discussion is the policy communication network [10]. Social network analysis (SNA) delivers clear and compelling visual interpretations of the intricate relationships within networks [11], providing valuable insights into the relationships, roles, and network structures engaged in the dynamic exchanges of interactions on social media in the realm of policy communication. In particular, ERGMs are extensively utilized to represent the stochastic process and the local structural characteristics underlying social networks, and as a powerful tool for analyzing and understanding the formation and evolution of complex networks [12]. The ability to comprehensively consider both network structure and node attributes [13] is crucial for revealing changes in the communication network structure before and after policy adjustments, thereby identifying shifts in the policy audience response. Communities within the policy communication network have played an important role in promoting policy dissemination and implementation through information sharing and resource exchange. However, their development still faces many challenges, such as the transparency and inclusiveness of the policy-making process [14].

To this end, the objective of this study is to address the following research questions: (i) how to analyze the structural characteristics of policy communication networks on social media using ERGMs; (ii) how to analyze the community characteristics of policy communication networks; and (iii) what changes occur before and after policy adjustments. Additionally, we conduct a case study of the Mass Entrepreneurship and Innovation policy, which has been instrumental in promoting innovation and entrepreneurship in China in recent years.

2. Literature Review

2.1. Policy Network

Policy networks are a concept first introduced by Katzenstein [15] as broadly defined organizational relationships, known as policy actor relationships, such as the relationships between the state, the social sector, and social alliances, which are political integrative structures formed during the policy-making process. Benson [16] defined policy networks from a sociological perspective as a complex of organizations that are interrelated due to their interdependence on resources. Rhodes and Marsh [17] believed that the structural relationships between political groups are the core of policy networks. Dowding [18] argued that policy network theory should be based on analysis of the network’s characteristics, rather than the characteristics of network members. Scholars have begun to analyze policy networks and construct models based on the social network analysis approach, promoting quantitative research on policy networks. Based on this foundation, Rhodes [19] further studied the defining characteristics of policy networks: trust, diplomacy, and reciprocity. He believes that the advantage of policy networks is their independence from the government, while the difficulties in policy network governance stem from how to manage diversity, the problem of many participants, the challenge of coordination, and local autonomy. Amiel et al. [20] analyzed policy networks in Israel’s education sector, demonstrating a transition from the traditional top-down model to a multi-stakeholder policy network approach. Policy network research from a stakeholder perspective has primarily focused on explaining the success or failure of policy efforts through resource dependence relationships [21]; this perspective provides an incomplete understanding of the social dynamics in policymaking.

2.2. Exponential Random Graph Models in Policy Research

Social network analysis provides valuable insights into the structure, dynamics, and patterns of relationships within policy networks, which can influence policy outcomes and effectiveness within a policy context [22]. SNA can examine the structural characteristics of social networks, such as network density, centrality, and connections between subgroups [23]. As one of the important research methods in social network analysis, ERGMs serve as an advanced statistical method that provides researchers with the opportunity to gain a deeper understanding of the structure and evolution of policy networks [24]. This approach possesses the prowess to approximate a maximum likelihood estimator for a given network dataset, leveraging Markov chain Monte Carlo simulations to generate new network datasets from a well-fitted ERGM [25]. ERGMs allow researchers to simulate and analyze both local structures and an entire structure within policy network data [26]. In policy research, ERGMs are used to assess how factors such as cooperative relationships between policy participants, information flow, and homogeneity of beliefs affect the formation and evolution of policy networks [27]. Jenkins-Smith et al. [28] used ERGMs to describe how policy actors form groups based on shared policy beliefs. By examining these local features, an ERGM infers the generative process of the entire network. Lee [29] studied the strategic collaborative relationships between local jurisdictions in a competitive environment and found that the physical, political, and geographical similarities between municipal authorities improve cooperative relationships between local jurisdictions, and that municipal authorities are more inclined to establish closed triangular relationships. Wagner et al. [30] proposed that ERGMs can reveal how endogenous and exogenous factors jointly influence the formation and change in network structures within policy networks. Endogenous factors include structural configurations within the network, such as reciprocity and triadic closure, whereas exogenous factors pertain to individual attributes and the institutional environment outside the network. Howe et al. [31] found that microstructural processes such as reciprocity, structural equivalence, and transitive closure in policy networks are crucial for explaining cooperative relationships within these networks.

2.3. Policy Communication Network Communities

Early research focused on the static structure of policy networks. However, with the widespread adoption of social media and digital tools, the focus has shifted to the dynamics, emotional mobilization, and mechanisms of information diffusion in communication networks. Social media platforms have made online communities an important arena for policy communication, not only providing new channels for the dissemination of policy information but also enhancing public participation and interactivity through the formation of networked communities [32]. The digital dissemination of policies on social media has driven the restructuring of collective actions that rely on policy network resources [33]. The formation of policy communication network communities can accelerate the rapid spread of policy information. Emotional mobilization within these communities, especially by opinion leaders, can drive the formation and evolution of overall community sentiment [34]. Policy communication network communities on social media exhibit distinct community characteristics, with users forming different communities based on interests, geographical locations, or political inclinations [35]. In the context of social media, the process of policy communication has seen the emergence of a new logic of “connective action” based on personalized content and framing. This logic emphasizes interactions between individuals and networked mobilization, rather than the traditional logic of collective action that relies on organizational intermediaries and collective identity [36]. These communities play a crucial role in the dissemination of policy information. Mayger et al. [37] conducted a Twitter-based analysis of the structural characteristics of policy information networks surrounding community schools in the United States. Their research revealed that distinct network communities formed around local policy agendas, demonstrating clear geographic differentiation in community structures.

Current scholarship predominantly utilizes either ERGMs or community analysis to examine the cooperative relationships in policy networks. Existing research on policy communication networks rarely adopts integrated methodologies for analyzing social media data, as the complexity of these platforms renders single-method approaches insufficient for comprehensive investigation and interpretation. In addition, comparative studies on network structures before and after policy modifications are particularly rare. Therefore, this study employs ERGMs and community analysis to investigate policy communication networks based on social media data, comparing the structural characteristics of communication networks before and after policy adjustments. These findings will offer valuable insights and guidance for policymakers in devising effective policy communication strategies.

3. Methods

3.1. Theoretical Framework

3.1.1. Policy Networks Based on Social Media

Although scholars have studied policy network actors and their relationships from the perspective of social network analysis, public policy is not solely made by a few core actors in dominant positions. In the policy process, there are also many individual actors and the small, interconnected groups they form, and the connections between policy actors and these individual actors or organizations influence policy outcomes [38]. Policy audiences, as the targets of policy communication, are not only recipients of policy information but also direct participants in policy implementation, an indispensable part of the entire policy process. Their actions, reactions, and interactions have a profound impact on the formulation, implementation, and evaluation of policies.

Social media has provided a platform that enables policy audiences to participate in the policy process in unprecedented ways [39]. In this new environment, policy networks are no longer limited to interactions between government agencies, interest groups, and experts but also include the voices and influence of the broad policy audience. The interactive nature of these platforms allows for immediate public response and participation, which is crucial for the iterative development of policies that are highly consistent with social needs [40]. Policy audiences, through interaction on social media platforms, can not only influence policy formulation but also supervise policy implementation and provide feedback on policy outcomes. The enhancement of this participation reflects the multidimensionality and dynamism of policy networks. Social media makes the boundaries of policy networks more blurred, the participants more diverse, and the interactions more frequent.

3.1.2. The Formation of Social Media Policy Communication Networks

The primary way for individuals to participate in social activities, groups are comprised systems or collections composed of individuals with special attributes or purposes. Under the influence of these close interactive relationships, information dissemination within groups is prone to form significant cohesion and consensus. Bennett and Iyengar [41] believe that the behavior of group communication on social media not only changes the way information is disseminated but also changes the way it is understood and the response to it. Social media group communication has become a key force in information dissemination and the formation of public opinion. A characteristic of social media group communication is decentralization [42]. On social media, any user can become a producer and disseminator of information. Social media users tend to interact with people who have similar views, looking for resonance. This not only satisfies the individual’s need for a sense of identity but also promotes the formation of social media group communication behavior.

The formation of social media policy communication networks is a multifaceted and highly interactive process, in which both policy communicators and policy audiences play important roles. Policy communicators clarify and publish policy information or directives through official channels, while media interpret these policies and disseminate them to a broader policy audience, becoming the initial nodes of the social media policy communication network because they possess policy information resources. Subsequently, the participation of opinion leaders and influencers on social platforms promotes the secondary dissemination of policy information, enhancing its impact. Under this influence, policy audiences engage in interactions regarding policy content, and their reactions and feedback become crucial, providing a feedback loop for policy makers to improve policies. During the interactive process, policy audiences become new nodes in the policy communication network, and the connections established between policy audiences and communicators, among policy audiences, and even among communicators themselves due to interaction, build the relationships between nodes in the policy communication network, forming numerous communities. The establishment of these interactive relationships and communities can promote and accelerate the retransmission and diffusion of policy information on social media platforms, contributing to the overall structure and expansion of the policy communication network.

Based on this, we propose the theoretical framework of this study. Building on the concept of policy networks and using social media as a platform, we incorporate the policy audience into the policy network to study the interactive relationship between communicators and the policy audience and construct a social media policy communication network. Furthermore, we employ social network analysis methods, particularly ERGMs, to analyze the characteristics of the social media policy communication network and communities. Figure 1 illustrates the theoretical framework for this study.

3.2. Social Network Analysis with ERGMs

As a statistical method for analyzing and simulating social network data, the ERGM is fundamentally based on the dissection of local structural characteristics within a network. The generative process of the entire network is inferred based on these local features. Within the ERGM framework, a network is considered a collection of a series of local structures, which include connections between nodes (edges), the number of connections a node has with other nodes (degree), reciprocity between nodes, and triangular structures.

In this study, the vertices represent the entities participating in policy discourse, whereas the edges symbolize the interactions among these entities. We designated a social media user ID as a vertex that qualifies as a participant in the policy communication network if they are engaged in such interactions. Consequently, any user IDs that do not partake in these interactions were naturally excluded from the network. Relationships that manifested as edges connecting social media users emerged from these interactions. Comments, shares, and likes constitute the fundamental modes of direct interaction between users. To delineate the research parameters, we established a boundary for this policy communication network, ensuring that all vertices and edges are predicated on policy-related interactions.

3.2.1. Local Network Patterns

We analyzed policy communication networks on social media by examining patterns within local network structures. A local network structure is a potentially small subgraph within the larger network fabric encapsulating localized regularities. As depicted in Figure 2a, the bidirectional arrows between the two nodes signify a mutual relationship, where both actors engage in interactive behaviors, such as liking, retweeting, or commenting on each other’s social media posts. In Figure 2b,c, ternary structures are commonly employed to denote local structures based on expansion (or convergence) and transitivity. Expansion is characterized by a local structure in which one actor extends relationships to a broader set of actors (left side of Figure 2b), whereas convergence is indicated when multiple actors direct their relationships towards a single actor (right side of Figure 2b). Transitivity refers to the formation of a unidirectional relationship between two actors through the intermediary of a third actor (node 1 points to node 2 and node 2 points to node 3). However, it is also possible that a direct relationship exists between the two actors, as illustrated in Figure 2c.

The emergence of an interactive relationship is contingent on the presence of two actors (nodes) rather than a solitary entity. Consequently, the focus of local network structures is on the dyadic connections and interactions between these nodes, particularly those that revolve around policy-related discourses. Although ternary structures add complexity by introducing a third node into the mix, they simultaneously amplify the potential for varied relationships, effectively layering multiple dyadic ties upon one another. Despite greater complexity, ternary relationships remain fundamental to local network analysis due to unique attributes like expansiveness, convergence, and transitivity, absent in dyads. These attributes remain instrumental in understanding network dynamics, even when the network expands to include additional nodes. Local network structures provide a detailed account of the minimal units of policy audience nodes and the fabric of the interactions that bind them. This foundational understanding is the bedrock of our approach to dissecting the architecture of policy communication networks. Ascertaining the presence of particular local network patterns within a specific policy communication network requires further empirical investigation.

3.2.2. Entire Network Structure Characteristics Analysis with ERGMs

The foundational concept of ERGMs is to conceptualize the genesis of a network as an aggregation of various local network structures, each reflective of distinct network structural traits. A composite network structure emerges as a complex interplay among these multiple local configurations. As discussed previously, the layering of nodes and edges leads to a compound superposition of local network outcomes. Within the process of overall network formation, there is a multi-faceted establishment of relationships and an intricate nesting of local network structures. As shown in Figure 3, the network encompasses various local network structures. The interplay and superposition of these localized configurations collectively delineate the comprehensive network architecture of patent collaborations within the seed industry for a specified region. These local structures are both probabilistically shaped and collectively transform the overall network, revealing its macro-level attributes. The likelihood of their presence and their collective impact are modulated by the model parameters.

This research utilizes the Markov Chain Monte Carlo (MCMC) approach to conduct the analysis. We explored the posterior distribution of the model parameters and identified parameter values that best matched the simulated network’s local structure to the observed network. This process aims to maximize the likelihood function of the observed network data, that is, the probability of observing the current network given the model parameters.

Specifically, for the ERGM of a network, the likelihood function $L$ can be expressed as:

$$L\left(\theta |y\right)=exp\left({\theta }^{T}g\left(y\right)\right)∕k\left(\theta \right)$$

• In which,

• $\theta$ is the vector of model parameters,
• $y$ represents the observed network,
• $g\left(y\right)$ is the vector of statistics for the network $y$ (such as the number of edges and triangles),
• ${\theta }^T$ denotes the transpose of $\theta$,
• $k$(θ) is the normalization constant that ensures the total probability of the model is 1.

The MCMC method establishes a Markov chain with a stationary distribution that aligns with the posterior distribution of the model’s parameters. In each iteration, according to the Metropolis–Hastings algorithm, we calculated the acceptance ratio $A$ for the new value ${\theta }^{\prime }$ of the parameter $\theta$:

$$A=min \left(1, L\left(\theta |y\right)k\left({\theta }^{\prime }\right)L\left({\theta }^{\prime }|y\right)k\left(\theta \right)\right)$$

If the acceptance ratio $A$ is greater than the random number $r$ drawn from the uniform distribution U(0, 1), then the new parameter value ${\theta }^{\prime }$ is accepted; otherwise, $\theta$ remains unchanged. In this manner, MCMC gradually explores the parameter space and converges to the optimal parameter configuration, thereby maximizing the probability of the observed network data.

The process of maximum likelihood estimation was used to fit the optimal parameters for edges, reciprocity, and triangles in a network model as $A_1$, $A_2$, and $A_3$. The network model was formulated as follows:

$$Mode{l}_{y }=A_1\ast edges+A_2 \ast mutual+A_3 \ast triangle+\epsilon$$

3.3. Community Analysis

Community refers to a subset of nodes within a network. The Number of Communities indicates the total number of communities or groups into which the network is divided, which can be detected using algorithms such as the Louvain method or the Girvan–Newman algorithm. The Description Length of Communities is a metric used to measure the complexity of a community structure, based on the concept of information theory, representing the amount of information required to describe the community structure in bits. It calculates the expected internal and external community connections, converting this into the information (bits) needed to describe the structure. A longer description implies a more complex community structure, which may include more diverse internal connection patterns or interactions between communities.

To thoroughly explore these communities and their characteristics, this study employs a community detection method based on modularity optimization. The Louvain algorithm initially treats each node as an independent community. It then iteratively optimizes modularity through two phases: modularity improvement and community aggregation, thereby uncovering high-quality community structures. In the modularity improvement phase, the algorithm adjusts the community affiliation of nodes by locally optimizing modularity. In the community aggregation phase, the already-formed communities are regarded as new supernodes to construct a new network, which continues to be optimized.

Modularity. Modularity is an indicator used to assess the quality of community partitioning. It is defined as:

$$Q={\displaystyle \frac{1}{2m}}\sum _{ij}\left[A_{ij }-{\displaystyle \frac{k_i{k}_j}{2m}}\right]\delta \left(c_i, c_j\right)$$

• In which,

• $A_{ij}$ is the weight of the edge between nodes $i$ and $j$,
• $k_i$ and $k_j$ are respectively the degrees of nodes $i$ and $j$,
• $m$ is the total weight of all edges,
• $c_i$ and $c_j$ are the communities to which nodes $i$ and $j$ belong,
• $\delta \left(c_i, c_j\right)$ is an indicator function, which equals 1 when $c_{i }$= $c_j$, and 0 otherwise.

Modularity increment. When a node $v$ moves from community $C_v$ to community $C^{\prime }$, the change in modularity is given by:

$$△Q=\sum _{u\in N\left(v\right)}\left[\delta \left(C_v , C^{\prime }\right)-{\displaystyle \frac{k_u{k}_v}{2m}}\right]$$

• In which,

• $N\left(v\right)$ is the set of neighbors of node $v$,
• $k_u$ is the degree of the neighboring node $u$.

The iterative process involves moving each node $v$ to the community of one of its neighbors, such that the modularity increment is maximized. Nodes belonging to the same community are then merged into a single supernode, and the edge weights, as well as the weights of edges between communities, are updated. These steps are repeated until the modularity no longer increases significantly or until a predetermined number of iterations is reached.

Sentiment Score. By analyzing the emotional orientation of communities, we can understand the attitudes and levels of support that individual communities have towards policy information. In terms of sentiment analysis, emotions are categorized into positive, neutral, and negative sentiments. On social media platforms, text analysis methods are employed to analyze the comments made by audiences around policy topics. In this study, a Naive Bayes classifier is used to achieve text sentiment classification. Within communities, the impact of each user node in the dissemination of policy information varies. Owing to factors such as social status, professional background, and the scale and quality of social networks, different user nodes have significantly different levels of influence over other users. For example, opinion leaders or well-known figures may have greater influence. Their views and attitudes can quickly spread and affect more people. Their emotional orientation and attitudes towards policy information can also directly or indirectly affect the dissemination methods and outcomes of policy information. If they hold positive sentiments towards policy information, it may promote positive dissemination of the policy and facilitate its implementation. Conversely, if they hold negative sentiments, it may lead to negative dissemination and hinder the smooth implementation of the policy. Therefore, in this study, the sentiment index is represented by the product of the audience node’s sentiment score and its number of followers.

Assuming $e_i$ is the sentiment score of node $i$ and $f_i$ is its number of followers, $S_i$ the weighted sentiment score of the community to which node $i$ belongs is shown in the formula as follows:

$$S_i={\displaystyle \sum }_{i=1}^n{e}_i\ast f_i$$

3.4. Mass Entrepreneurship and Innovation Policy and Policy Adjustment

China’s economy is undergoing a strategic metamorphosis, shifting from an investment-centric growth model towards one that is innovation-led and driven by creative prowess. The government has spearheaded a nationwide initiative to catalyze innovation and entrepreneurship across the country—the Mass Entrepreneurship and Innovation policy. In contrast to other state-implemented policies, this policy boasts a more extensive reach, encompassing a diverse array of stakeholders such as enterprises, universities, research institutions, and various organizations, alongside individuals eager to embrace innovation and embark on entrepreneurial ventures [43]. The MEI policy comprises a policy system that is constantly adjusted and optimized. In June 2015, the State Council issued Opinions of the State Council on Several Policies and Measures to Vigorously Promote Mass Entrepreneurship and Innovation (hereinafter referred to as Version 1.0). Three years hence, China’s economy transitioned from a phase of swift expansion to one characterized by high-quality development. As this evolution precipitated novel demands for fostering widespread entrepreneurship and innovation, the Opinions of the State Council on Promoting the High Quality Development of Innovation and Entrepreneurship was issued in 2018, creating an upgraded MEI policy (hereinafter referred to as Version 2.0). The commentary acknowledges that the wave of mass entrepreneurship and innovation has surged with remarkable success. However, it also candidly addresses ongoing issues. In response, the state is visibly making concerted efforts to recalibrate policies, aiming to extend and deepen the reach and impact of mass entrepreneurship and innovation initiatives.

3.5. Data and Processing

To trace the trajectory of policy evolution, we procured two iterations of policy documents from the official government portal of the People’s Republic of China. In our quest to assess the communicative impact of the MEI policy on social media, we selected the Sina Weibo platform as our primary repository for gauging public reactions. The data used in our case study were exclusively sourced from publicly available user-generated content and public comments on the Sina Weibo platform. Sina Weibo facilitates a thematic discussion feature, allowing policy audiences to congregate and articulate their perspectives and concerns within dedicated policy discussion threads. Given that the gap between the initial (Version 1.0) and subsequent (Version 2.0) policy drafts spanned three years and three months, we endeavored to mitigate the temporal influence on the efficacy of policy communication. Accordingly, we conducted a comparative analysis of the policy communication effects within the same timeframe. To analyze the responses, we selected data within the same period (i.e., one year for each) after each version of the policy was introduced (Figure 1, $t_1$ for Version 1.0, $t_2$ for Version 2.0). We then separately collected data on policy topics in $t_1$ and $t_2$ (Figure 4).

We conducted a targeted search for data, employing the keywords “mass innovation and entrepreneurship” prefixed with the topic symbol, to streamline our retrieval process. The term “mass innovation and entrepreneurship” serves as the policy’s official moniker, encapsulating the most recurrent themes within the document’s lexicon. In the Chinese context, these phrases operate as two distinct yet complementary concepts. Additionally, the Chinese abbreviation for the policy can take on various forms, including the insertion of a space, period, or slash to denote separation in the middle. Occasionally, users may even invert the order of the two terms. The peculiarity is the policy’s adoption of a two-character Chinese abbreviation, which, when rendered into English, is succinctly captured as “Double Creation”. In other words, several topics need to be searched to obtain the policy discussion data. Thereafter, manual screening was performed to exclude repeated posts, advertisements, or commercial information, and posts irrelevant to the policy.

4. Results

Figure 5 illustrates the network structure of “mass entrepreneurship and innovation” policy dissemination during two time periods, $t_1$ and $t_2$, before and after the policy adjustment. It can be visually observed that before the policy adjustment, the network formed for policy dissemination had more nodes and more complex interactions. Conversely, after the policy adjustment, although some more prominent communities emerged within the dissemination network, there were relatively fewer nodes, and the interactions were more dispersed.

4.1. ERGM Analysis Results

We employed R software (R 4.4.2) to conduct an ERGM analysis of the communication network characteristics before and after the adjustment of the MEI policy. $Mode{l}_1$ is the fitted model for the policy communication network at time $t_1$, and $Mode{l}_2$ is the model for time $t_2$. The results returned by the R software showed that both models underwent two iterations, and there were differences in the parameter estimates of the policy communication network in the ERGM.

Therefore, $Mode{l}_1$ and $Mode{l}_2$ are as follows:

$$Mode{l}_1=-8.029516\ast edges+2.803711\ast mutual+2.966579\ast triangle$$

$$Mode{l}_2=-7.898816\ast edges+1.289087\ast mutual+12.875506\ast triangle$$

• Edges: In both models, the parameter estimates for the edges were negative, indicating that edge formation in the network was suppressed. However, the negative estimate was more pronounced in $Mode{l}_1$ (−8.029516 vs. −7.898816), suggesting a more significant suppression of edges.
• Mutual: The parameter estimate for reciprocity in $Mode{l}_1$ was positive (2.803711), indicating that reciprocal connections were facilitated in the network. Conversely, the estimate for $Mode{l}_2$ was lower (1.289087), suggesting that reciprocity had a less noticeable effect in $Mode{l}_2$.
• Triangle: The parameter estimate for triangular structures in $Mode{l}_1$ was positive (2.966579), indicating that triangular structures were promoted in the network. The estimate for $Mode{l}_2$ was significantly higher (12.875506), indicating a more pronounced presence of closed triads (i.e., triangles) in the network represented by $Mode{l}_2$.

First, the negative parameter estimates for edges in both networks before and after policy adjustment suggest that the number of connections (edges) tends to be lower than expected in a random network. This implies that certain social structures or communication mechanisms inhibit actual connections in policy communication networks. A smaller edge parameter estimate after policy adjustment indicates a more pronounced inhibitory effect on the network, which may reflect more selective connections or obstacles in the policy communication process.

Second, the positive parameter estimate for reciprocity before policy adjustment indicates that if one node (such as an individual or organization) disseminates policy information to another, the latter is also inclined to disseminate information, indicating strong mutual communication and interactions. By contrast, the lower value of the reciprocity parameter after the policy adjustment suggests that reciprocal connections are not a major driving force in the network. This may mean that policy information dissemination in the post-adjustment network is more unidirectional, or that reciprocal connections are not a significant feature in policy communication.

Finally, the positive parameter estimate for triangular structures before policy adjustment indicates a tendency for nodes to form closed triads, which may foster close connections and concentrated information dissemination within subgroups or communities. A significantly higher triangular parameter value in the dissemination network structure following policy adjustment indicates that such triangular structures are highly prevalent in the network, which may indicate the presence of many tight-knit subgroups or communities that play a core role in policy communication.

4.2. Communities Analysis Results

4.2.1. Community Overview

In terms of community numbers, before the policy adjustment, there were 231 communities in the communication network (

Table 1. Comparison of Community Characteristics in the Communication Network before and After Policy Adjustment.

	Before	After
Number of Clusters	231	214
Modularity	0.944773	0.808050
Maximum Number of Levels	4	3

). After the policy adjustment, the number of communities decreased slightly to 214. The reduction in community numbers may be due to smaller communities merging into larger ones under the guidance of the policy, or some marginal communities gradually disappearing as a result of the policy adjustment. However, overall, the impact of the policy adjustment on the number of communities is not significant.

Regarding modularity, the modularity of the communication network was 0.944773 before the policy adjustment, and it decreased to 0.808050 after the adjustment. Before the policy adjustment, the elements within communities in the policy network were strongly connected, with high internal cohesion among members who were tightly linked by common goals and policy topics, ensuring the smooth circulation of policy information. After the policy adjustment, the original community structure changed, and the internal tightness of communities was reduced.

In terms of hierarchical levels within communities, before the policy adjustment, the maximum number of hierarchical levels for policy information dissemination within a single community was four. After the adjustment, this number decreased to three. The simplification of hierarchical levels in the communication network’s communities after the policy adjustment may indicate that the process of policy information dissemination has changed, allowing for faster and more direct dissemination. It could also suggest that some barriers between hierarchical levels were broken down, simplifying the previously multi-layered and nested dissemination structure.

4.2.2. Community Distribution

Before the policy adjustment, the community distribution in the communication network was relatively dense, with many communities (Figure 6). The entire network exhibited a complex and diverse structure, with different communities interwoven with each other. The larger communities had numerous and intricate connections, indicating frequent information exchange and interaction between communities. The channels for policy information dissemination were diverse, and each community could easily exchange policy information with others.

After the policy adjustment, the community distribution in the communication network became relatively sparse, with a noticeable decrease in the number of large communities (Figure 7). Although connections between communities still existed, the interactive relationships between them were significantly reduced compared to before the adjustment. Some communities became more centralized, meaning they were more likely to form simple but relatively large communities centered around a particular node. This may indicate that after the policy adjustment, the types of community members became somewhat homogenized, with certain types of members representing a higher proportion in specific communities. This suggests that the pathways for policy information dissemination were simplified after the policy adjustment, with some previously complex channels of information exchange being integrated or severed. However, it may also lead to situations where marginal communities have insufficient access to or exchange of policy information.

4.2.3. Community Sentiment Analysis

In this study, the sentiment orientation of communities is set within the range of [−100, 100]. Negative values indicate that user nodes hold negative attitudes towards the policy, 0 indicates a neutral or wait-and-see attitude towards the policy, and positive values indicate a positive attitude towards the policy. The closer the value is to −100 or 100, the stronger the negative or positive attitude of the user nodes towards the policy. The sentiment index also considers the number of followers for each node. Nodes with many followers may influence their followers to form positive attitudes if their comments are positive. Conversely, if their comments are negative, they are likely to influence their followers to hold negative attitudes towards the policy.

Looking at the overall sentiment orientation of the top 10 communities, significant changes have occurred. Both before and after the policy adjustment, the main communities maintained positive sentiment orientations (

Table 2. Sentiment Analysis of Major Communities before and after Policy Adjustment.

Before				After
Key Nodes	Number of Nodes	Average Sentiment Score	Weighted Sentiment Score	Key Nodes	Number of Nodes	Average Sentiment Score	Weighted Sentiment Score
Xindu News, Jintang Release	198	46.46	69.00	National Mass Entrepreneurship and Innovation Activity Week—Beijing Venue	350	58.40	71.95
Live Sichuan	148	58.41	66.45	Harbin Release	336	30.58	47.16
Micro View of Nanyang	107	10.16	20.12	Selected News Highlights	229	49.74	70.88
Youth of Fuzhou University	103	57.49	64.38	Zhejiang Release	226	55.61	63.97
Class 1 of Grade 13, Mathematics and Computer Science, Fuzhou University	99	39.70	63.48	Zhao Bosi Jobs	199	65.61	67.03
FLAT Entrepreneurship	83	52.14	55.88	Tonglu Release	143	69.44	73.23
Youth League School, Fuzhou University	82	58.77	60.21	AliOS	103	45.32	78.34
Environmental and Safety Engineering Student Union, Fuzhou University	70	52.26	55.23	Sparse Mountain Foliage in Fine Weather for Fruit Picking	94	17.23	12.61
Class 2 of Grade 14, Material Control, Fuzhou University	69	54.81	63.97	Beijing Release	67	48.83	65.84
Yulin Propaganda	66	39.44	66.53	Creativity is Simple	55	49.43	74.17

). Before the policy adjustment, the weighted sentiment scores of the main communities ranged from 20.12 to 69.00. After the policy adjustment, the weighted sentiment scores ranged from 12.61 to 78.34. This indicates that although the main communities remained positive in sentiment after the policy adjustment, a polarization in sentiment orientation emerged.

Both before and after the policy adjustment, there were communities with relatively low sentiment orientations within the main communities. Analysis of comments from community members revealed that their key nodes and some members, while discussing the MEI policy, also reflected certain negative social phenomena. These phenomena may not be directly related to the MEI policy, but the key nodes believed that such phenomena did not reflect the entrepreneurial and innovative spirit that the policy aimed to promote. Instead, they felt that these phenomena hindered the implementation of the MEI policy and were contrary to its original intent.

5. Discussion

This study analyzed changes in the characteristics of the communication network structure on social media and changes in interaction before and after an MEI policy adjustment, using ERGMs and community analysis. The findings were as follows.

After the policy adjustment, the number of triangles in the communication network structure significantly increased, but the formation of edges remained constrained.

The communication network exhibited a predominantly positive triangular configuration both prior to and following the policy adjustments. After the policy adjustment, the triangular interaction structure increased significantly compared to before the policy adjustment. This indicates that user nodes in the policy communication network are more inclined to form closed triads, meaning that the connections between user nodes are not just bilateral, but have begun to form closer and more stable three-way relationships. New incentive mechanisms brought about by the policy adjustment may have contributed to this phenomenon. The MEI policy adjustment resulted in the expansion of the scope of support for both individual innovators and entrepreneurs, with a greater focus on applying specific measures to promote innovation in key directions. Additionally, the responsible departments and institutions are listed for each support measure, which is conducive to forming a consensus on the development goals of innovation and entrepreneurship across society. All these factors can stimulate more interactions on policy topics and promote closer network connections.

However, the negative parameter estimates from the ERGM results indicate that the formation of new interactive relationships in the communication network faced certain obstacles or restrictions both before and after the adjustment of the MEI policy, similar to the research results of Howe et al. [31]. The nature of the data suggests that public attention to policy is relatively low on social media compared to contemporary events, and the scale of user groups participating in policy interaction is relatively small. In other words, users on social media platforms pay more attention to the content or events they are interested in compared to policy information, with relatively low preferences for information selection, which inhibits the establishment of interactive relationships in policy topics [44]. While policies tend to receive more attention when first released, as time goes by, despite adjustments in content, there may still be a decay effect in information dissemination, with a decrease in attention [45].

After the policy adjustment, cross-community connections in the communication network decreased, and communities exhibited localized contraction.

Structural Level. Before the policy adjustment, the communication network had an umbrella-shaped radiating structure, which transitioned to a relatively independent and closed structure after the adjustment. The number of cross-community connections significantly decreased, with interactions and exchanges becoming more concentrated within individual communities. This shift may be attributed to the evolution of the policy and the transition from mobilization-based participation to vertical management in its implementation mechanism. As innovation and entrepreneurship policies moved from macro-level advocacy to specific implementation, the pathways for information dissemination inevitably underwent organizational restructuring along administrative lines. This led to the localization of cross-regional discussions into vertical channels.

Actor Attribute Level. The primary community participants shifted from “campus innovators and entrepreneurs” to “government services”. Before the policy adjustment, college students, as one of the main beneficiary groups of innovation and entrepreneurship policies, actively engaged in policy discussions under the impetus of campus “innovation and entrepreneurship” activities, forming multiple interactive communities centered around college students. After the policy adjustment, local governments became more prominent, with official government microblogs in various regions emerging as the main leaders of these communities. This outcome corroborates the “government intervention effect in communication structure” proposed by Bennett and Segerberg [33]. The reason is that local governments became the core force in policy implementation and promotion. Governments in regions such as Beijing, Zhejiang, and Harbin introduced specific policy measures to foster innovation and entrepreneurship. Through interactions involving relevant implementation departments, they promoted policy publicity and strengthened the effectiveness of policy implementation. This transformation reflects the inevitable evolution of the policy life cycle from the advocacy stage to the implementation stage.

After the policy adjustment, although communities overall maintained a supportive stance, a trend of polarization emerged.

Before and after the policy adjustment, the main communities exhibited positive sentiment orientations, indicating support. However, after the policy adjustment, the range of sentiment fluctuations within these communities widened. In other words, considering the number of followers, some communities developed more positive sentiments, while others showed negative inclinations. The positions and attitudes of key nodes played a decisive role in the overall sentiment direction of their respective communities. If key nodes held negative attitudes towards the policy, the weighted sentiment orientation of their communities might also lean towards negativity. This suggests that in the process of policy dissemination, the influence of key nodes cannot be ignored, as their attitudes and behaviors can significantly impact the overall sentiment orientation of communities.

While previous studies have often attributed sentiment polarization to political stance differences [46], this research finds that the polarization of community sentiment is closely related to policy communication strategies. After the policy adjustment, the community centered around the key node “Harbin Release”, despite including numerous nodes such as government agencies, media, and celebrities, exhibited a relatively low level of sentiment. The primary reason was the mechanical communication approach adopted by these nodes, which simply transplanted policy content in the form of press releases onto social media, lacking commentary, guidance, or encouraging language. This rigid and stereotyped communication method failed to resonate emotionally with the public, preventing policy messages from deeply engaging the audience and potentially leading to cognitive biases [47]. Emotional resonance is crucial in policy communication, as it can bridge the gap between policies and the public, stimulating a sense of identification and participation. If policy communication fails to touch the emotions of the public, its influence and implementation effectiveness will inevitably be significantly reduced. Therefore, key nodes should focus on in-depth content processing and emotional guidance when disseminating policies to enhance their communication effectiveness and impact.

6. Conclusions

To analyze the evolutionary patterns of policy interaction, this study employs ERGMs and community analysis to simulate the structure of policy communication networks. The theoretical contributions are made in two aspects. The first is the extension of the application scenario of policy network theory to social media, proposing the concept of policy communication networks based on policy network theory, and studying and analyzing policy communication on social media through social network analysis methods. The second is the inclusion of policy audiences as important actors in the policy communication network, which expands the scope of actors in the policy network. These policy audiences are not only recipients of policy information but also practitioners of the policy interaction.

The findings of this study hold significant implications for management practices. First, regarding Structural Intervention in Policy Communication Networks, to address the issue of “inhibited edge formation”, a multi-dimensional online interaction incentive mechanism can be established. For example, interactive point systems and policy participation certifications could be introduced to encourage users to earn points or certifications through activities such as sharing, in-depth commenting, or creating original policy interpretations. This would attract policy audiences to engage in interactive exchanges. Additionally, role embedding can be used to promote cross-regional dialogues. For instance, virtual policy liaison roles can be created to initiate cross-regional policy discussions, alleviating localized closures. This approach not only strengthens the vertical depth of local policy implementation but also activates the potential for policy innovation through cross-border connections, ultimately achieving a structural leap from “closed collaboration” to “open co-governance” in communication networks. Second, shifting from Reactive Public Sentiment Management to Proactive Emotional Governance, to mitigate the risk of emotional polarization among key nodes in policy communication, a key node emotional early warning system can be established utilizing AI-based sentiment analysis tools to monitor sentiment dynamics in real-time, such as introducing multimodal sentiment detection algorithms (text/image/video), setting negative sentiment monitoring thresholds for communities and opinion leaders, and establishing an intervention feedback mechanism. When the negative sentiment index exceeds the threshold, a tiered intervention mechanism will be automatically triggered. When negative emotions are detected among key nodes, timely emotional counseling can be provided. Moreover, key nodes that actively disseminate positive content can be given traffic support or rewards, forming a positive cycle of “emotional guidance—influence enhancement”. Through these mechanisms, a paradigm shift from reactive public sentiment management to proactive emotional governance can be achieved, continuously optimizing the emotional ecology of policy communication networks. Third, to enhance Policy Communication Strategies through Emotional Resonance, instead of mechanical content dissemination, policy communication should focus on emotional guidance and resonance. Policy content should be deeply processed and transformed into multi-dimensional expressions with emotional impact. This can be achieved by incorporating specific examples and storytelling techniques to enhance the readability and attractiveness of policies. Emotional resonance can bridge the gap between policies and audiences, using scenario-based narratives to evoke empathy and stimulate a sense of identification and participation.

This study also has profound implications for political and social dimensions. First, the implementation of structural interventions and emotional governance mechanisms can effectively enhance policy transparency and public participation, thereby strengthening government credibility and improving governance efficacy. Second, the proposed cross-regional dialogue mechanisms and emotional early-warning systems help mitigate local protectionism and regional development disparities, facilitating resource sharing and policy innovation, and contributing to coordinated regional development. Finally, proactive emotional governance strategies can effectively address social polarization, reduce conflicts arising from emotional opposition, and foster social harmony and stability, creating a favorable environment for the smooth implementation of policies.

Future studies could employ ERGMs for predictive and intervention analyses of policy communication networks. By integrating ERGM network analysis with sentiment computing technologies, the structural evolution and emotional fluctuations of policy communication networks can be tracked, providing data support for strategy adjustments.

However, this study has some limitations. First, it does not show the entire dynamic evolution of the MEI policies. In addition, because users can change their privacy rights on Sina Weibo, some data can no longer be obtained. Further studies should follow up with long-term policy evaluations using real-time data acquisition.