A Review of Group Polarization Research from a Dynamics Perspective

Abstract

The rapid rise of social media has accelerated the evolution of public opinion, leading to frequent group polarization. Meanwhile, advancements in information science have enabled large-scale experiments, positioning dynamics as a crucial perspective for studying group polarization. This paper systematically reviews group polarization from a dynamics perspective. First, we outline its definitions and its explanatory theories. Then, we examine the role of dynamics in polarization research, summarize the current measurement methods of group polarization, and analyze intervention strategies based on elements of dynamics. Finally, we propose a logical framework for dynamics-based interventions. Our findings indicate that while research on group polarization from a dynamics perspective is relatively comprehensive, most intervention studies remain at the simulation level, requiring further validation for real-world applicability. This review provides a systematic overview of group polarization through a dynamics lens, offering insights for addressing challenges in network governance within the social media era.

1. Introduction

Group polarization refers to the social phenomenon where the positions of group members become more extreme after engaging in discussions on a particular issue (Isenberg, 1986). It has been one of the significant research projects in the social sciences in recent decades. In many instances, group polarization is considered a product of political ideologies and party affiliations (Yarchi et al., 2020). During group discussions, individual political orientations (Robison & Moskowitz, 2019), regional backgrounds (Fershtman, 1997), and the evolving context of events (Schmidt et al., 2018) are closely intertwined with their speech and behavior (J. Jiang et al., 2020). Individuals often have specific thoughts and opinions based on their original ideologies in response to particular events or topics. However, group polarization is not a simple aggregation of individual opinions, but the mutual reinforcement of each member’s stance during interaction, causing the group’s opinion to shift to extremes. The reason is that within a group, individuals often lack rational judgment and are highly prone to being influenced by the opinions of others (Le Bon, 2018). Wallach et al. (1964) have found that when a group shares responsibility, it makes individuals more susceptible to the influence of other opinions, causing individual behavior to tend toward extremism. Lee (2006) further has found that the occurrence of group polarization usually requires three conditions: a platform for group discussion, members’ freedom to express their opinions, and the group reaching a certain size.

With the application and development of internet technologies, group polarization in virtual worlds has begun to attract widespread attention. American psychologist Wallace (1997) proposed that group polarization in online environments is more pronounced than in the real world, as it is easier to find individuals with similar opinions on a specific topic within the internet space. Today, the convenience of the internet facilitates the rapid spread of public opinion, with new media (social media) gradually replacing traditional media as the primary channel for information dissemination and the main platform for public discourse, thus becoming a key platform for observing group polarization (Yuan et al., 2022). Group polarization on social media can be explained from two perspectives: the individual perspective and the platform perspective. From the individual perspective, users are exposed to vast amounts of information, and while they may encounter posts containing opposing opinions, selective communication leads users to prefer receiving and sharing content that aligns with their pre-existing opinions (Gillani et al., 2018). This behavior can be explained by the concept of the information cocoon, which posits that in the process of information dissemination, due to individuals’ limited information needs, people tend to focus only on the communication domains they choose and that bring them comfort, gradually placing themselves in a cocoon (Sunstein, 2006). The information cocoon emphasizes the subjective psychological choices of individuals, with Sunstein (2001) arguing that the information cocoon created by like-minded individuals is a key mechanism of group polarization. From the platform perspective, recommendation algorithms lead platforms to recommend content that reinforces users’ existing beliefs and social identities, filtering out information that users are not interested in (Spohr, 2017; Xing et al., 2024). This personalized information environment created for internet users by personalized filters is known as the filter bubble. Filter bubbles can impede users from encountering heterogeneous information, underscoring the role of algorithms in clustering similar users through personalized information services (Pariser, 2011). The information cocoon and the filter bubble limit the diversity of information through individual selective psychology and platform technological means, restricting the scope of information reception for users and thus leading to the homogenization of information. This homogenization manifests itself in a relatively closed environment where voices with similar opinions are repeated in exaggerated or otherwise distorted form, causing the majority of people in that environment to believe that these distortions are the whole truth. This phenomenon is the echo chamber, also known as the stratospheric effect (Sunstein, 2001; Bessi et al., 2015; Mocanu et al., 2015; Yilmaz et al., 2024). Hong and Kim (2016), Schmidt et al. (2018), Brady et al. (2020), Cinelli et al. (2021), Van Bavel et al. (2021), and Hobolt et al. (2023) argued that social media platforms lead individuals into echo chambers and stimulate conflicts among groups, thereby exacerbating group polarization. Gentzkow and Shapiro (2011), Prior (2013), and Boxell et al. (2017) suggested that social media does not necessarily create echo chambers in discussions, nor does it play a role in group polarization. However, Semaan et al. (2014) and Beam et al. (2018) indicated that social media usage might reduce group polarization. Regarding the role of social media in group polarization, Iandoli et al. (2021) and Arora et al. (2022) conducted review studies. However, social media has indeed exacerbated group polarization in many cases, leading to various negative social impacts, such as the spread of misinformation (Tucker et al., 2018), the exacerbation of intergroup conflict (Iyengar et al., 2019), and declines in social cohesion (Kingzette et al., 2021).

In order to systematically review the relevant literature, several scholars have conducted comprehensive reviews from different perspectives. Isenberg (1986) reviewed the research on group polarization between 1974 and 1982, sorted out the two main explanatory theories of group polarization, namely social comparison theory and persuasive argumentation theory. He concluded that the combination of the two theories produces group polarization, and persuasive argumentation theory plays a greater role. Following the emergence of social media as a primary platform for group polarization, Iandoli et al. (2021), through an examination of 121 research papers on group polarization in social media, summarized the antecedents, theoretical explanations (mechanisms), enabling design affordances, and effects of online group polarization. Their findings confirmed that social media exacerbates group polarization by amplifying offline social processes. Furthermore, Arora et al. (2022) conducted a systematic review of research on social media and group polarization, identifying the contingencies and mechanisms of social media’s impact on group polarization. These mechanisms include five key elements: information processing and sharing, socio-psychological processes, civic and political engagement, network and platform drivers, and content and framing aspects. They proposed a conceptual framework for understanding the multifaceted impacts of social media on group polarization, arguing that existing research presents contradictory opinions on whether social media exacerbates or mitigates group polarization, and there is a need to move beyond the dichotomy of social media as beneficial or harmful. The above studies have systematically summarized the phenomenon of group polarization and its mechanisms in the context of social media, providing a theoretical foundation for understanding the formation process and internal mechanisms of group polarization.

However, while existing studies generally suggest that the negative impacts of group polarization dominate (Iandoli et al., 2021), effective intervention methods have yet to be systematically identified. Moreover, when determining appropriate intervention timing and evaluating the effectiveness of such interventions, it becomes necessary to quantify group polarization. With the widespread use of social media and the rapid development of information technology and computer science, research on group polarization from a dynamics perspective has made significant progress. Therefore, this paper will conduct a systematic review of the phenomenon of group polarization from a dynamics perspective, proposing effective intervention methods and frameworks.

Building on existing theoretical and empirical research on group polarization, this study aims to achieve the following objectives:

• Conduct a systematic review of the definitions and explanatory theories of group polarization.
• Analyze the dynamics perspective in group polarization research, focusing on three key aspects: propagation dynamics, opinion dynamics, and the coupled dynamics of propagation and opinion.
• Summarize current methods for measuring group polarization.
• Organize existing intervention strategies based on the elements of dynamics and propose a logical framework for dynamics-based interventions.

By systematically reviewing group polarization from a dynamics perspective, this study seeks to provide a comprehensive theoretical reference for researchers in the field and offer research-driven insights for addressing network governance challenges in the evolving social media landscape.

2. Methodology

This study follows the guidelines for systematic literature review (SLR) proposed by Kitchenham and Charters (2007), which is important in the field of knowledge to improve the quality of research (Gonzalez Camacho & Alves-Souza, 2018). These guidelines include three main stages of structured reviews, which are (1) planning the review, (2) running the review, and (3) reporting the findings. The following is a detailed description of each stage.

2.1. Planning the Review

This study aims to answer the following questions:

RQ1: How can group polarization be studied through kinetic modeling?

RQ2: What are the methods for measuring group polarization?

RQ3: How can group polarization be mitigated?

The database search engines selected for this study include Scopus, Web of Science, EBSCOhost, and IEEE Xplore. In order to obtain literature that matches the topic, the keywords are identified as “group polarization”, “dynamics”, “propagation dynamics”, and “opinion dynamics”. The study of selected rules is limited to English.

2.2. Running the Review

To ensure the quality of the literature, the study is performed using abstracts, using the following syntax (“group polarization”) AND (“dynamics” OR “propagation dynamics” OR “opinion dynamics”). The search resulted 222 results from Scopus, 97 results from Web of Science, 24 results from EBSCOhost, and 55 results from IEEE Xplore. Following duplicate removal, a total of 306 articles are retained for subsequent analysis.

2.3. Reporting the Review

In this study, 306 articles mentioned above are reviewed through a two-step process. The first step is an initial review that assesses the relevance of the article to the topic and its potential to address the three issues mentioned above based on its title, abstract, and keywords. A total of 227 articles passed the initial review. The second step is a textual review, in which articles that passed the initial review are further examined, and final inclusion is decided based on the content of their full text. A total of 143 articles eventually passed the text check, forming the dataset for this study.

3. Definition and Explanatory Theory

3.1. Definition of Group Polarization

The theory of group polarization originated in the field of social psychology. Stoner (1961) found that when individuals made decisions as members of a group, they tended to adopt more risky courses of action than when making decisions as individuals, the phenomenon known as risky shift. Then, Nordhoy (1962) found that group discussions could also lead to more cautious decisions, proposing the concept of cautious shift, which suggests that some decisions become more conservative and cautious after discussion. Subsequent studies further supported either risky shift or cautious shift (Brown, 1965). Stoner (1968) concluded that for items whose values tend to favor risky choices, the average of group decisions tends to be riskier than the individual’s initial decision, while for items whose values tend to favor cautious choices, group decisions tend to be more cautious. Pruitt (1971) argued that the emergence of risky shift was due to researchers framing decision problems in terms of risk, whereas the emergence of cautious shift resulted from framing problems in terms of caution. He proposed that the more appropriate term for both risky shift and cautious shift was choice shift. The group polarization was initially described as choice shift. Moscovici and Zavalloni (1969) referred to both risky shift and cautious shift as the polarization effect of group interaction and suggested that polarization effects were not confined to the domain of risk but could occur in any area of group interaction. However, they also noted that polarization effects do not always occur. When the object of discussion is of little significance or unfamiliar to the majority of members, individuals will not express strong preferences, and thus, the polarization effect will not occur. Fraser et al. (1971) were the first to combine the terms group and polarization, suggesting that the term group polarization more accurately described the phenomenon of choice shift. The concept of group polarization gradually became concrete over time. Myers and Lamm (1976) defined group polarization as the tendency for the average postgroup response to become more extreme in the same direction as the average pregroup response. Isenberg (1986) defined group polarization as the tendency for group members to adopt more extreme opinions on a given issue after engaging in discussion, and this definition has been widely used by scholars from the perspective of choice shift.

As research has progressed, scholars have increasingly defined group polarization from the perspective of multi-group opposition, suggesting that it refers to the state or tendency of a group to split into two or more conflicting or sharply contrasting subgroups (Evans, 2003; Martín-Gutíerrez et al., 2023). With advancements in dynamics and network science, the conceptualization of group polarization has become more precise. Esteban and Ray (1994) and Fiorina and Abrams (2008) characterized it as a bimodal distribution of opinions among group members. Building on this, Lelkes (2016) further argued that group polarization manifests as the intensification of this bimodal distribution. Gross and De Dreu (2019) described the transition from numerous small communities to a few large ones as a process of group polarization, with the highest level of polarization occurring when a group consolidates into two dominant opposing communities.

These studies suggest that the early definitions of group polarization from the perspective of choice shift mainly described the phenomenon of unipolar polarization, while the later definitions from the perspective of multi-group opposition mainly described the phenomena of bipolar and multipolar polarization.

3.2. Explanatory Theory of Group Polarization

Scholars in sociology and psychology have proposed several explanatory theories to account for group polarization, including persuasive argument theory, social comparison theory, social identity theory, social norm theory, self-categorization theory, and information influence theory. Each of these theories provides a distinct perspective on the mechanisms driving group polarization, as summarized in

Table 1. Explanatory theory of group polarization.

Theory	Core Assumption	Explanation of Group Polarization	Representative Literature Support
Persuasive argument theory	Individuals are more susceptible to more persuasive arguments to change their opinions, and the persuasiveness of the argument is determined by validity or acceptability and novelty.	Studies have shown that when individuals exhibit a bias toward a particular direction, they tend to provide additional arguments supporting this direction, thereby reinforcing the initial inclination, which leads to group polarization.	(Vinokur, 1969; Burnstein et al., 1971; Vinokur & Burnstein, 1978; Kaplan, 1977; Sia et al., 2002; Fu & Zhang, 2016)
Social comparison theory	Individuals evaluate themselves by comparing themselves with others and then adjust their behavior to meet society’s expectations of the individual.	Studies have shown that individuals compare their opinions with those of others and adjust their opinions to be more in line with the prevailing direction, expressing their opinions more strongly when they find the same opinions, which leads to group polarization.	(Festinger, 1954; Brown, 1965; Sanders & Baron, 1977; Sia et al., 2002)
Social identity theory	Individuals develop specific attitudes and behaviors based on their social identity to enhance a sense of identity and belonging.	Studies have shown that individuals tend to accept and reinforce the dominant opinions of the group in order to maintain and enhance group identity, suggesting that the formation of group identity is driven by socialization rather than rational thinking, which leads to group polarization.	(Jost et al., 2003; Huddy et al., 2015; Törnberg et al., 2021; Wuestenenk et al., 2023)
Social norm theory	Social norms are the common expectations of group members regarding behavior, attitudes, and values. Individual behavior is constrained by these social norms.	Research has shown that individuals fear social isolation, so they have to change their opinions to match those of the group, which leads to group polarization.	(Glynn & Noelle-Neumann, 1986)
Self-categorization theory	Individuals categorize themselves into appropriate groups based on the similarities between self and group and adjust their opinions to fit the characteristics of the group to which they belong.	Studies have shown that individuals tend to adjust their opinions to align with those of their reference group, with some individuals categorizing themselves as belonging to more extreme factions within the group, which leads to group polarization.	(Abrams et al., 1990; McGarty et al., 1992; Hogg & Reid, 2006)
Information influence theory	Individuals facing uncertainty rely on surrounding information for their decisions and behaviors, especially when the information is coherent.	Research has shown that individuals tend to accept and disseminate the viewpoints within the group that are information-rich in order to quickly acquire information and make decisions, thereby intensifying homogeneity, which leads to group polarization.	(Myers & Lamm, 1976)

.

Isenberg (1986), through a review of studies on group polarization conducted prior to 1982, identified social comparison theory and persuasive argument theory as the two predominant explanatory frameworks. He found that while both theories contribute to the occurrence of group polarization, the effects of persuasive argument theory are often more pronounced. Empirical research by Butler and Crino (1992) further supported this view, suggesting that social comparison and persuasive argument processes can occur simultaneously during group discussions, thereby leading to group polarization. This implies that, in real-world contexts, group polarization is likely shaped by multiple underlying mechanisms. Aloka and Bojuwoye (2013) reinforced this perspective in their study on disciplinary hearing decisions, demonstrating that both persuasive arguments and social comparisons significantly influence polarization outcomes. Additionally, Iandoli et al. (2021) argued that social comparison theory and social identity theory work together to fulfill the need for internal cohesion, while persuasive argument theory facilitates group survival and adaptation by confronting external threats. These findings suggest that group polarization cannot be explained by a single mechanism alone; rather, multiple explanatory processes may coexist and interact within a single instance of polarization.

These theories explain the formation of group polarization from sociological and psychological perspectives, highlighting key factors such as individual psychology, information dissemination, and the social environment. They provide valuable research perspectives and analytical frameworks for a deeper understanding of the dynamics underlying group polarization, thereby promoting the advancement of dynamics-based research in this field. Additionally, these insights offer a theoretical foundation for developing effective intervention strategies to address group polarization in real-world scenarios.

4. Dynamics Models of Group Polarization

Although the above explanatory theories reveal the multiple factors and mechanisms underlying group polarization, they are often limited to empirical analysis and are typically framed at either the macro (group-level) or micro (individual-level) scale. Consequently, they struggle to capture the emergence of group polarization as a micro-to-macro process. Dynamics offers a novel research approach to address this limitation. As a method for studying complex systems, it enables the quantification of group polarization elements and facilitates an in-depth exploration of its underlying mechanisms. Furthermore, the widespread use of social media, alongside advancements in information technology and computer science, has provided both extensive data and the computational infrastructure necessary for conducting large-scale computational experiments within a dynamics-based research framework. As a result, dynamics has progressively become a dominant perspective in the study of group polarization.

Currently, research on group polarization from a dynamics perspective primarily focuses on two key areas: propagation dynamics and opinion dynamics. Additionally, some scholars have explored coupled dynamics models, which integrate both propagation and opinion dynamics to provide a more comprehensive understanding of polarization processes.

4.1. Propagation Dynamics Model

Propagation dynamics is a crucial area of research in both communication studies and complex network theory, focusing on the processes and patterns of propagation in fields such as infectious diseases (Keeling et al., 2008), behavior (Centola, 2018), and information dissemination (W. Wang et al., 2019). Propagation dynamics primarily concerns the extent of transmission, the rate of spread and outbreak thresholds. The models of propagation dynamics are primarily divided into two categories: infectious disease models and influence propagation models. Infectious disease models include the classical Susceptible-Infected (SI) model (Pugliese, 1990), Susceptible-Infected-Susceptible (SIS) model (Kermack & McKendrick, 1933), and Susceptible-Infected-Recovered (SIR) model (Kermack & McKendrick, 1927), along with several variant models such as Susceptible-Infected-Recovered-Susceptible (SIRS) model (Cheng et al., 2017), Susceptible-Exposed-Infected-Recovered (SEIR) model (Prem et al., 2020), Susceptible-Vaccinees-Infected-Susceptible (SVIS) model (Shim, 2006), and Susceptible-Vaccinees-Infected-Recovered (SVIR) model (Liu et al., 2008). Influence propagation models mainly include the linear threshold (LT) model (Schelling, 1971) and the independent cascade (IC) model (Banerjee, 1992).

Because of the similarities between information dissemination and infectious disease transmission, researchers have extensively applied infectious disease models to study information dissemination (Clifford & Sudbury, 1973). Among the above models, the infectious disease models and the LT model are usually integrated with opinion dynamics models, which will be discussed in Section 4.3 of this paper. Regarding the IC model, Del Vicario et al. (2016) developed a data-driven percolation model to simulate rumor spreading. They proposed that homogeneity is the primary driving force behind content diffusion, with polarization clusters being a common outcome, and indicated that homogeneity and polarization are key determinants for predicting the size of cascades. Dai et al. (2022) established a link between individual preferences and information, creating a realistic dynamic enhanced independent cascade (EIC) model. To maximize positive influence while minimizing negative influence, the authors defined the effective influence maximization with the group polarization (EIMGP) problem and proposed a group-based effective influence maximization (GEIM) algorithm that combines metaheuristic strategies. This algorithm is capable of distinguishing between positive and negative opinions, thereby identifying higher-quality seed nodes. These studies reveal the formation mechanism of group polarization from the perspective of propagation dynamics, mainly by simulating the information dissemination process.

4.2. Opinion Dynamics Model

Opinion dynamics is a crucial aspect of sociophysics, exploring the processes and patterns of group opinion interactions across various fields such as marketing (J. Chen et al., 2021), public opinion management (Y. Dong et al., 2018; Bashari & Akbarzadeh-T, 2019), political elections (Di & Galanis, 2021), and collective decision-making (J. Dong et al., 2024). Opinion dynamics primarily focuses on three stable states of opinion: consensus, bipolarization, and fragmentation (Y. Dong et al., 2018; B. Jiang et al., 2023), which correspond to unipolarization, bipolarization, and multipolarization. Based on the form of opinion representation, opinion dynamics models can be classified into two main categories: discrete and continuous opinion dynamics models. In discrete opinion dynamics, individual opinions are binary, including the voter model (Holley & Liggett, 1975), majority rule model (Galam, 2002), and Sznajd model (Sznajd-Weron & Sznajd, 2000). In continuous opinion dynamics, individual opinions lie within a continuous range, including averaging models such as the DeGroot model (DeGroot, 1974) and the Friedkin-Johnsen (FJ) model (Friedkin & Johnsen, 1997), as well as bounded confidence models such as the Deffuant-Weisbuch (DW) model (Deffuant et al., 2000), the Hegselmann-Krause (HK) model (Hegselmann & Krause, 2002), and the Jager-Amblard (JA) model (Jager & Amblard, 2005).

In the models above, averaging models and bounded confidence models are often applied in the study of group polarization. Kooshkaki et al. (2023) developed partisan belief opinion dynamics models based on the DeGroot model, namely the Partisan Confidence-lite (PC-lite) and Partisan Confidence (PC) models and derived the conditions under which group polarization occurs in both models. Musco et al. (2018) formalized the minimization of polarization as a convex optimization problem based on the FJ model, optimizing the objective function in terms of changing both the network structure and the initial opinions to mitigate group polarization, and applied the method to Twitter and Reddit, finding that the level of group polarization was significantly reduced. B. Jiang et al. (2023) proposed a set of inclusiveness functions that integrate conflict and compromise mechanisms based on the DW model. They constructed an inclusiveness-degree based signed Deffuant-Weisbush (ISDW) model and analyzed the evolution of opinions under different conditions of bounded confidence and inclusiveness-degree. Their findings showed that under conditions of low inclusiveness and high confidence, opinion bipolarization rapidly transitions into unipolarization. Xu et al. (2023) proposed an opinion-climate-based HK dynamics model, in which the opinion climate was quantified as the relative proportion of individuals holding positive or negative opinions and incorporated as a factor in opinion interaction rules. Their findings demonstrated that under the influence of opinion climate, individual opinions would eventually reach a consensus. C. Wang et al. (2019) incorporated information distortion and polarization effects into an information entropy-based opinion dynamics model to simulate the uncertainty in human memory and communication. They found that large-scale polarization toward positive or negative consensus occurs when the synergistic mechanism between preferential trust and polarization tendencies persists. This study effectively analyzes and predicts opinion polarization phenomena on social platforms. These studies reveal the formation mechanism of group polarization from the perspective of opinion dynamics, mainly by simulating the opinion interaction process.

4.3. Coupled Dynamics Models

Research on group polarization from the perspective of propagation dynamics mainly focuses on the dissemination mechanism of information, which is unidirectional dissemination, and generally does not involve the interaction of various kinds of information and usually focuses on the communication process of information in groups at a small-time scale (Li et al., 2021). In contrast, research on group polarization from the perspective of opinion dynamics mainly focuses on the interaction mechanisms of opinion, which is multi-directional, and needs to study the interaction between opinions, and usually focuses on the interaction mechanism of public opinion (including values, trends, etc.) on a larger time scale (Y. Dong et al., 2018), as shown in Figure 1. In many real-world scenarios, such as the eruption of collective events, the formation of social trends, or the popularity of products, the processes of information dissemination and opinion interaction do not occur independently. While information is being disseminated, individual opinions simultaneously interact. Consequently, some scholars have begun to integrate propagation dynamics with opinion dynamics, studying the group polarization through the coupled mechanism of both processes.

Current research on the coupled models of propagation dynamics and opinion dynamics generally follows the approach of determining changes in individual states through propagation dynamics models and quantifying individual opinion values through opinion dynamics models. Li et al. (2021) defined and quantified factors influencing information dissemination based on the SEIR model and the HK model, such as user influence, topic popularity, and topic interest. They proposed the public opinion evolution HK-SEIR model and found that when the density of infected users is high, multiple homogeneous groups emerge within the network. While group polarization is a typical manifestation of network homogeneity, they suggested that information dissemination is more likely to induce group polarization. Yuan et al. (2022) developed an SIR model considering dynamic network structures based on the SIR model and the JA model. They investigated the effects of dynamic network adjustments, infection rate, and immunization rate on bipolarization and validated the model’s rationality and effectiveness. T. Chen et al. (2020) established a group polarization model integrated into the SIRS epidemic model based on the SIRS model and the JA model. They studied the effects of the immune recovery rate and relationship strength on group polarization and found that the BA network is more bipolarized than the small-world network on the same scale. Yin et al. (2023) constructed the SLFI-JA model based on the Susceptible-Latent-Forwarding-Immune (SLFI) model and the JA model, considering both internal (individual-level) and external (environmental-level) factors. By comparing normally distributed opinions with bipolar and unipolar opinion distributions, they discovered that individuals holding strong support or opposition opinions facilitate information diffusion. Wu et al. (2023) proposed a depolarization model based on the modified LT model and HK model, and used the model to design three depolarization strategies, which were verified by simulation experiments. These studies reveal the formation mechanism of group polarization from the perspective of the coupling between propagation dynamics and opinion dynamics, mainly by simulating the information dissemination and opinion interaction process, fully considering both information dissemination and opinion interaction in the process of public opinion evolution.

Combining the above three research perspectives, we found that the research models of group polarization under different research perspectives are highly consistent. First, the model is constructed through the mechanism of message dissemination or opinion interaction, and then the laws of public opinion dissemination or interaction are studied based on the model. Some studies will also explore in depth effective intervention measures for group polarization.

5. Measurement

In current research, the degree of group polarization is often quantified after constructing a model to analyze its patterns. Scholars have developed various measurement approaches to assess polarization levels.

To present these measurement methods clearly, we first define key variables. Let $n$ denote the number of individuals or agents, and let the set of opinions be represented as $X=\left\{x_1,x_2,\cdots {,x}_n\right\}$, where $\overline{X}$ represents the mean opinion of the group, and $P$ denotes the group polarization index. For continuous opinions, opinion values range from [−1,1], where −1 represents a negative opinion and 1 represents a positive opinion.

A review of the literature reveals that when measuring group polarization in discrete opinion models, scholars typically do not assign specific numerical values to opinions but instead focus on categorical distinctions.

5.1. Choice Shift

From the perspective of choice shift, the degree of group polarization is defined as the difference between everyone’s opinion $x_i\left(0\right)$ prior to discussion and the final group opinion $\overline{X\left(T\right)}$ (Sia et al., 2002):

$$P={\displaystyle \frac{1}{n}}\left|{\sum }_{i=1}^n(\overline{X\left(T\right)}-x_i(0\left)\right)\right|$$

(1)

5.2. Multi-Group Opposition

Current research mainly measures group polarization from the perspective of multi-group opposition. Based on this, we categorize the measurement methods of group polarization under the multi-group opposition framework according to the form of opinions. Most measurement approaches are focused on bipolarization phenomena.

• Discrete opinion

Starting from the social network behavior of individual users, group polarization can be measured by observing individual preferences (Schmidt et al., 2018). For topics with two opposing opinions (bipolarization issues), assuming that user $i$ likes viewpoint $X_1$ for $m$ times and likes viewpoint $X_2$ for $n$ times, the polarization level of user $i$ is defined as Equation (2):

$$p_i={\displaystyle \frac{(m-n)}{(m+n)}}$$

(2)

The value of $p_i$ approaching −1 indicates that user $i$ is polarized towards viewpoint $X_2$, while a value approaching 1 indicates that user $i$ is polarized towards viewpoint $X_1$. The degree of group polarization is then reflected by the probability density function of the polarization levels $p_i$ across all users. However, this approach has certain limitations, as it is challenging to accurately quantify the extent of group polarization based solely on the final state of the presented probability density function.

Starting from the distribution of group opinions, group polarization can be measured using the index for qualitative variation (IQV) (Agresti & Agresti, 1978; Buskens et al., 2008). Let $P\left(X_1\right)$ denote the proportion of individuals in the group who support opinion $X_1$, and $1-P\left(X_1\right)$ represent the proportion of individuals who support opinion $X_2$. The IQV is calculated as Equation (3):

$$IQV=4P\left(X_1\right)(1-P(X_1\left)\right)$$

(3)

When the value of $P\left(X_1\right)$ approaches 0.5, and the value of IQV approaches 1, indicating that the group tends towards bipolarization. Conversely, when the value of $P\left(X_1\right)$ approaches 0 or 1, and the value of IQV approaches 0, indicating that the group tends towards unipolarization (consensus).

2.

Continuous opinion

When individual opinion is continuous, the degree of group polarization can be measured using the bimodality coefficient (BC) of the opinion distribution function $f\left(X\right)$ (Lelkes, 2016; Musco et al., 2021). The calculation of the bimodality coefficient is as Equation (4):

$$BC={\displaystyle \frac{{\gamma }^2+1}{\kappa +\frac{{3(n-1)}^2}{(n-2)(n-3)}}}$$

(4)

$\gamma$ refers to the skewness of the distribution, and $\kappa$ refers to its kurtosis. $BC=1$ indicates that the distribution is perfectly bimodal, while $BC=0$ indicates that the distribution is perfectly unimodal. Additionally, when $BC$ exceeds $1/3$, the opinion distribution is considered to be bipolarized.

Inspired by the concept of electric dipole moment, some researchers propose that individuals with differing opinions are analogous to two oppositely charged point charges, with the degree of polarization depending on the extent of conflict between their opinions, that is, the distance between the two opinions (Fershtman, 1997; Brum et al., 2022). Let $p\left(x\right)$ represent the probability density distribution of individual opinion values, with $X^{+}$ and $X^{-}$ defined as the relative totals of positive opinions ($x>0$) and negative opinions ($x<0$) respectively. $d^{+}$ and $d^{-}$ are defined as the centroids of positive and negative opinions. By calculating the opinion difference $∆X$ and the normalized distance $d$, the group polarization index can be defined as Equation (7):

$$∆X=\left|X^{+}-X^{-}\right|=\left|P(x>0)-P(x<0)\right|=\left|{\int }_0^{1}p\left(x\right)dx-{\int }_{-1}^{0}p\left(x\right)dx\right|$$

(5)

$$d={\displaystyle \frac{\left|d^{+}-d^{-}\right|}{\left|X_{max}-X_{min}\right|}}={\displaystyle \frac{\left|\frac{{\int }_0^{1}p\left(x\right)xdx}{{\int }_0^{1}p\left(x\right)dx}-\frac{{\int }_{-1}^{0}p\left(x\right)xdx}{{\int }_{-1}^{0}p\left(x\right)dx}\right|}{2}}$$

(6)

$$P=(1-∆X)d$$

(7)

When $P=1$, $∆X=0$ and $d=1$, the opinion distribution is fully bipolar, with half of the individual opinions concentrated at the two poles, −1 and 1. When $P=0$, $∆X=1$ or $d=0$, the opinion distribution is unipolar, concentrating at a single extreme point or at neutral opinions. However, this calculation method has a notable limitation, as it is difficult to precisely determine the state of the opinion distribution solely based on the value of the group polarization index $P$.

In the case of multipolarization, Azzimonti and Fernandes (2022) categorized individuals holding different opinions into distinct groups. From the perspective of the homogeneity of opinions within groups and the heterogeneity of opinions between groups, the group polarization index is defined as Equation (8):

$$P=C{\sum }_{k=1}^G{\sum }_{l=1}^G{\pi }_{k,t}^{1+\zeta }{\pi }_{l,t}\left|{\overline{x}}_{k,t}-{\overline{x}}_{l,t}\right|$$

(8)

$C$ is a constant greater than 0, $G$ represents the number of groups, and $\zeta$ is a positive parameter that reflects group identification (Esteban & Ray, 1994). ${\pi }_{k,t}$ denotes the proportion of individuals in group $k$ at time $t$ relative to the total population. ${\overline{x}}_{k,t}$ and ${\overline{x}}_{l,t}$ are the average opinion values of individuals in groups $k$ and $l$ at time $t$, reflecting the homogeneity of opinions within groups. $\left|{\overline{x}}_{k,t}-{\overline{x}}_{l,t}\right|$ reflects the heterogeneity of opinions between groups. The value of $P$ approaches 1 as the degree of group polarization intensifies, and approaches 0 as polarization weakens, indicating a more uniform distribution of individual opinions. This measurement method effectively captures the strength of group polarization. However, it requires prior knowledge of the number of groups, which may pose practical challenges in its application.

In the measurement of group polarization from the perspective of multi-group opposition based on continuous opinions, certain metrics do not differentiate between the specific number of groups, but instead measure polarization based on the divergence of individual opinions.

Based on the individual opinion values themselves, group polarization can be measured through the magnitude of the opinion vector $\mathit{x}$ (Matakos et al., 2017):

$$P={\displaystyle \frac{1}{n}}{\left|\left|\mathit{x}\right|\right|}^2$$

(9)

Based on the deviation of individual opinions from the mean, group polarization can be measured through the variance of opinion values (Musco et al., 2018; Brooks & Porter, 2020; Luskin et al., 2022; Yin et al., 2023):

$$P={\displaystyle \frac{1}{n}}{\sum }_{i=1}^n{(x_i-\overline{X})}^2$$

(10)

Based on the differences in individual opinions, group polarization can be measured using the global disagreement index (GDI) (Dandekar et al., 2013) and the network disagreement index (NDI) (Bindel et al., 2015; X. Chen et al., 2018). The NDI extends the GDI by incorporating network structure. Although this increases computational complexity, it significantly reduces the computational scale. In this context, $w_{ij}$ represents the strength of the influence of individual $i$ on individual $j$’s opinion, with $w_{ij}\ge 0$ and ${\sum }_{j=1}^n{w}_{ij}=1$.

$$GDI={\sum }_{i<j}{(x}_i-x_j{)}^2$$

(11)

$$NDI={\sum }_{(i,j)\in E}{w_{ij}(x}_i-x_j{)}^2$$

(12)

Spohr (2017) proposed that group polarization can be studied from both micro and macro perspectives. We observe that most measurement approaches originate from the micro level and ultimately establish the measurement methodology based on the macro level. In addition to the targeted indicators above, scholars have also employed community size, user collaboration frequency, social network distance (SND), network modularity, and the normalized cut between communities as measures to reflect the degree of group polarization (Amelkin et al., 2019; Flamino et al., 2023; Phillips et al., 2023; Iyer & Yoganarasimhan, 2021). In practical research, it is necessary to comprehensively consider the computational complexity and scale according to specific problems and available data types, make full use of data, and select appropriate quantification methods, so as to provide a reference for determining the intervention opportunity and evaluating the intervention effect of group polarization.

6. Impact and Intervention Measures

Group polarization can have both positive and negative impacts, depending on the specific issues at hand in various domains. In response to these impacts, interventions targeting group polarization may be considered. This section reviews the impacts of group polarization and outlines potential intervention measures based on the elements of dynamics models.

6.1. Impact of Group Polarization

Group polarization significantly influences multiple areas, including democratic politics and governance (Yuan et al., 2022), social stability and welfare (Tucker et al., 2018), and economic development (Q. Wang et al., 2012). It is often a direct consequence of groupthink. When groupthink fosters collective intelligence, groups may make more effective decisions than individuals, leading to positive outcomes from polarization (Aronson et al., 2017). However, group polarization also has negative consequences. Iandoli et al. (2021) categorized the detrimental effects of group polarization on social media into three key areas: fragmentation of the online public sphere, opinion radicalization, and the spread of online misinformation.

Our review of existing literature reveals that the impacts of group polarization primarily concern four main domains: politics, society, economy, and culture.

Table 2. Impact of group polarization in different areas.

Research Area	Research Theme	Positive Impact	Negative Impact
Politics	Political position	Increase individual trust in institutions and their representatives (Johnson et al., 2017).	Erupt into radicalism or civil war (Fershtman, 1997); Reduces people’s rationality significantly (Druckman et al., 2013); Worsen the relationship between the party and the people and intensify conflicts and divisions (Robison & Moskowitz, 2019; Brum et al., 2022).
	Public policy	NA	Damage the government’s image (Yuan et al., 2022); Reduce citizens’ trust and participation in the democratic process (Jones, 2015; McCoy & Somer, 2018).
	Ideology	NA	Individuals are manipulated by opinion leaders to mask self-awareness (Bekafigo et al., 2019); Reduce the diversity of perspectives in a democratic system (Benson, 2023).
Society	Public opinion dissemination	NA	Contribute to the spread of falsehoods and misinformation (Tucker et al., 2018; Vicario et al., 2019; Maia et al., 2023); Limit the influence of social media providing accurate information (Schmidt et al., 2018); Public discussions become disorderly and chaotic, and even lead to cyber violence (J. Jiang et al., 2020).
	Social stabilization	Increased focus on social issues and formation of cohesive groups on issues of consensus (smoking cessation, alcoholism, charitable giving, etc.) (El-Shinnawy & Vinze, 1998).	Exacerbate racial prejudice (Sunstein, 2001); Generates individual feelings of hatred and exclusion of potentially fair participants, affecting the normal order of society (X. Wang & Song, 2020); Increase inter-group conflict and decreased social cohesion (Iyengar et al., 2019; Kingzette et al., 2021).
Economics	Economic order	NA	Manipulate the financial market (Kiymaz, 2002; Spiegel et al., 2010).
	Company profits	Enhance customer enthusiasm and increase revenue (Luo et al., 2013)	Increase escalation of investment in failing business ventures (Brockner, 1992; Whyte, 1993); Negative reviews damage reputation and long-term profits (Dai et al., 2022).
Culture	Values	NA	Interfere with value judgments and value choices (Ohtsubo et al., 2002); Dissolve the moral consensus (Gonçalves-Segundo, 2022).
	Outcomes	Produce higher-quality articles and generate group wisdom (Shi et al., 2019).	NA

summarizes the positive and negative impacts of group polarization within these domains. Although some experimental studies have demonstrated the positive effects of group polarization, the negative impacts are more prevalent. Furthermore, Schmidt et al. (2018) and Falkenberg et al. (2024) have shown that the degree of polarization tends to increase over time. Given this trend, interventions are necessary to mitigate the negative consequences of group polarization.

6.2. Intervention for Group Polarization

Given the negative effects of group polarization, it is crucial to intervene in the process of public opinion evolution. Drawing from dynamics models, we review existing interventions for group polarization from three key perspectives: individual opinions, evolutionary mechanisms, and network structure. In the context of propagation dynamics, the evolutionary mechanism refers to the rules governing information dissemination, while in opinion dynamics, it pertains to the rules of opinion interaction (Y. Dong et al., 2018). Based on this framework, we propose a logical framework for dynamics-based interventions aimed at addressing group polarization.

6.2.1. Group Polarization Interventions Based on Individual Opinions

When intervening in group polarization through individual opinions, there are usually two approaches: one is to alter the opinion of the original individuals, and the other is to introduce new individuals, namely opinion leaders (Although this approach also changes the network structure, it is logically focused more on altering the opinions themselves).

In terms of altering the opinion of the original individuals, Matakos et al. (2017) defined two NP-hard problems in the FJ model, ModerateInternal and ModerateExpressed, based on the internal and expressed opinions of individuals. The objective is to identify a set of nodes and set the internal or expressed opinion values of the nodes in the set to 0, in order to minimize the group polarization index. They also designed polynomial-time algorithms to solve these two problems. Musco et al. (2018) employed a convex optimization program when modifying the agents’ innate opinions and solved it using standard algorithms in polynomial time to reduce the bipolarization.

In terms of introducing new individuals, Yin et al. (2023) introduced opinion leaders (the highest-degree nodes in the network) as cohesion disruptors and examined the effects of opinion intervention intensity and timing on the scale and peak of information diffusion. They found that after introducing opinion leaders, both the diffusion scale and peak decreased, with earlier intervention having a more significant impact, thus effectively controlling the spread of negative public opinion. Wu et al. (2024) introduced social bots to generate opinions and examined their depolarization effects, revealing that such bots could mimic rational users with positive viewpoints, effectively mitigating negative group polarization. Yuan et al. (2022) discovered that as the initial opinion value distribution gap of newly introduced individuals increased, individuals with more extreme attitudes gradually emerged, thereby significantly exacerbating bipolarization. When the initial opinion value distribution gap of new individuals was small, the group’s opinions tended to be more moderate.

The intervention measures based on individual opinions have diverse algorithms, but they may have certain limitations when applied to specific problems. Algorithms that alter the existing individual opinion values face challenges in scalability and computational efficiency, particularly in large-scale networks, and their operability remains to be further evaluated. Introducing new individuals presents the risk of reduced opinion diversity, which could affect the stability of group dynamics. For interventions based on individual opinions, future research could focus on the computational efficiency and adaptability of algorithms in complex network environments, while also reasonably determining the extent of introducing opinion leaders based on measures of group polarization.

6.2.2. Group Polarization Interventions Based on Evolutionary Mechanisms

When intervening in group polarization through evolutionary mechanisms, it is possible to approach the issue from two perspectives: altering the information dissemination rules from the perspective of propagation dynamics, or altering the opinion interaction rules from the perspective of opinion dynamics.

In terms of the information dissemination rules in propagation dynamics, Yuan et al. (2022) developed a network opinion polarization model based on the SIR model. Their study found that reducing the infection rate or increasing the immunization rate could mitigate the degree of bipolarization. The infection rate refers to the probability that information disseminators influence an individual and subsequently becomes a new disseminator. This rate is largely dependent on the level of attraction the event holds for the group. Therefore, it is crucial to timely reduce the public attention generated by group events to prevent extreme public opinion. The immunization rate refers to the probability that an individual ceases to participate in the spread of information due to forgetfulness or disagreement. Consequently, authoritative information should be disseminated through multiple channels to enhance individuals’ immunity to false or misleading information, thus preventing them from becoming disseminators of misinformation. T. Chen et al. (2020) built a group polarization model that is integrated into the SIRS epidemic model, and discovered that as the immune recovery rate increases, the bipolarization phenomenon becomes more pronounced. The immune recovery rate indicates the probability that an immune individual will be transformed into an information disseminator again after being stimulated by new, strong evidence. Therefore, it is necessary to effectively release authoritative information so that individuals can accurately identify the authenticity of the information.

In terms of the opinion interaction rules in opinion dynamics, B. Jiang et al. (2023) developed an ISDW model to investigate the effects of confidence level and inclusiveness degree on opinion formation. The confidence level refers to the range of opinion differences an individual is willing to accept, reflecting the likelihood of communication between individuals holding different opinions. The inclusiveness degree refers to the probability that an individual will accept the opinions of others within the range of their confidence level, reflecting the individual’s capacity to accept opinions differing from their own. The study found that as the confidence level increases, the number of opinion clusters decreases, and the opinion distribution transitions from fragmentation to bipolarization and ultimately to consensus, with a corresponding reduction in convergence time. When the inclusiveness degree increases, a higher confidence level is required for the formation of opinion clusters. Yin et al. (2023), through the establishment of the SLFI-JA model, discovered that with an increase in confidence level, the assimilation effect gradually dominates. In such cases, individuals tend to adopt more conservative opinions, and the group opinion gradually shifts toward neutrality.

The intervention measures based on evolutionary mechanisms make full use of elements from different dynamics models, altering factors such as the infection rate and immunization rate in the dissemination rules, as well as the confidence level in the interaction rules, thereby influencing the outcome of public opinion evolution. However, in practical applications, effectively altering these factors presents certain challenges, and further research is needed to develop specific implementation methods.

6.2.3. Group Polarization Interventions Based on Network Structure

When intervening in group polarization through network structure, most studies consider the changes in the edges, and a few studies consider the changes in the nodes.

Musco et al. (2018) mathematically formalized the problem of reducing overall disagreement as an optimization problem. They proposed an algorithm to determine the topology that minimizes opinion divergence, given the number of agents, their initial opinions, and the edge weights in the network structure. They demonstrated that there always exists a graph with edges of $O(n/{ϵ}^2)$, and that a network with very sparse connections can effectively minimize opinion divergence. Garimella et al. (2017) represented the discussion process of controversial issues as an endorsement graph and proposed an efficient algorithm for identifying edges that minimize the degree of bipolarization. This algorithm recommends opposing viewpoints, breaks filter bubbles, and considers the acceptance probability of recommended edges, which is more efficient than other algorithms.

The above research only considered changes in the edges between nodes, while some scholars have also examined the addition or removal of nodes. Yuan et al. (2022) found that under identical conditions, the degree of group polarization in static networks is significantly higher than that in dynamic networks. This is because the dynamic adjustment of the network encourages some extreme individuals to exit from the previously rigid network, while newly added non-extreme individuals can help alleviate the trend of bipolarization to some extent.

The intervention measures based on network structure are currently more applicable to group discussions on social media platforms. In practice, their implementation depends on the specific platform, offering good feasibility and efficiency. However, it is essential to carefully calibrate the scope of such interventions and uphold ethical standards when platforms intervene, to prevent the emergence of new issues while striving for the desired outcomes.

6.2.4. Logical Framework for Group Polarization Interventions

During group discussions, the timing of group polarization interventions and the evaluation of their effectiveness can be determined based on specific measures of group polarization. When implementing interventions, measures should be taken from elements of dynamics, namely individual opinions, evolutionary mechanisms, and network structure. The explanatory theories of group polarization informed by sociology and psychology should provide guidance, while feedback mechanisms are used to adjust the level of intervention. Ultimately, this process outputs the results of the group discussion. The logical framework for dynamics-based interventions in group polarization is illustrated in Figure 2.

In practical applications, during the initial stages of public opinion events, the polarization of online debates can be measured and monitored through specific group polarization metrics. When the polarization value reaches a certain threshold, a warning signal is triggered (Alsinet et al., 2021), prompting further intervention measures. Additionally, the transparency of platform algorithms also affects group polarization. Xing et al. (2024) summarized eight associations of group polarization based on previous literature, establishing the PolarSphere ecosystem driven by cognitive biases on social media. They proposed that transparent and responsible algorithm design can mitigate the amplification of biased content. Furthermore, the formulation of healthier online discourse regulations is necessary to maintain a delicate balance between freedom of speech and the reduction of harmful polarized content. However, current research on interventions for group polarization primarily relies on simulation experiments, and given the complex and dynamic nature of real-world scenarios, further empirical studies are required to evaluate the effectiveness of these interventions.

7. Conclusions and Discussion

This paper offers an extensive and in-depth analysis of the existing literature on group polarization from a dynamics perspective. First, we systematically review the definitions and explanatory theories of group polarization, elucidating the mechanisms underlying its formation. This provides an analytical framework and a logical reference for the study of group polarization through the lens of dynamics. Second, we survey representative research on group polarization within the dynamics framework, analyzing current studies across three key aspects: propagation dynamics, opinion dynamics, and the coupled dynamics of propagation and opinion. Furthermore, we organize and analyze various methods for measuring group polarization, offering insights into determining the optimal timing and assessing the effectiveness of interventions. Finally, by synthesizing the impacts of group polarization across different fields, we find that its negative effects dominate. Based on the elements of dynamics, including individual opinions, evolutionary mechanisms, and network structure, we organize and analyze the intervention measures found in current models of group polarization, proposing a logical framework for dynamics-based interventions. This systematic analysis of group polarization from a dynamics perspective aims to provide a comprehensive reference for scholars in the field.

In current research, scholars employ various methods to measure group polarization. This paper focuses on categorizing the different measurement approaches and then explores intervention strategies. We analyze and summarize the strengths, weaknesses, and applicable contexts of each measurement method and intervention. In practical applications, it is crucial to consider the types of data available in the given problem, alongside computational complexity, feasibility, and adaptability, to select an appropriate measurement approach and evaluate potential interventions. Moreover, interventions must be developed with reference to sociological and psychological explanatory theories of group polarization. However, most current research on interventions relies on simulation experiments. The applicability of these methods in internet governance remains an open question for future research. Empirical studies are needed to explore the effectiveness of these interventions. We hope that this paper will offer a scientific basis and actionable ideas for addressing the challenges of internet governance in the current digital landscape.