Network Effects Underlying Belief Polarization: An Agent-Based Modelling Approach

1. 群体中的信仰两极分化是全球政治（例如，Levendusky, 2010）和环境（Cook and Lewandowsky, 2016）事务的重要挑战。例如，尽管有可疑的论证和缺乏足够的支持证据，极化的党派精英仍会塑造公众信仰（Druckman， Peterson和Slothuus， 2013）。因此，研究两极分化与不稳定信念的形成之间的关系是对抗民主发展危害的关键（Flynn， Nyhan和Reifler， 2017）。社交媒体资源经常被用来寻找有关政治内容的信息（例如，Mitchell et al., 2014），最近的研究调查了社交媒体在信仰极化背景下的作用（例如，Flaxman， Goel和Rao， 2013）。研究结果表明，反复暴露于选择性选择的信息（例如，通过Facebook的算法排名的News Feed）减少了与先前信念不一致的信息消费（Bakshy， Messing和Adamic， 2016）。总的来说，这些发现强调了理解信念极化机制的重要性。理解这些机制可以帮助政策制定者和系统架构师调解信仰两极分化的出现。此外，它将允许更多地了解与形成分歧的极端主义信仰相关的潜在利益和合理性（Levendusky, 2010; Jern， Chang和Kemp， 2014）。2. 当前提案的重点是最近使用基于主体的模型（ABMs）和迭代学习的工作（例如，Griffiths和Kalish, 2007; Kirby， Griffiths和Smith, 2014； Cornish等人，2017;Madsen， Bailey和Pilditch， 2018）。特别有趣的是Madsen等人（2018）最近发表的一篇论文。在他们基于主体的模拟中，作者研究了网络结构如何在假设所有主体都是理性的情况下（在按照贝叶斯定理形成信念的意义上）导致回声室（ECs）的形成。令人惊讶的是，研究发现，尽管没有认知和社会偏见，ECs还是出现了。事实上，在非常简单的假设下，理性主体之间的相互作用可以产生ec，其严重程度与网络规模呈正相关。这一发现对有关群体信仰极化及其后果的文献综述具有重要意义。首先，它强调了足以导致ec形成的最小结构方面（例如，代理之间的相互作用和网络大小）。其次，它表明，为了减少极端主义信仰的固定，政策制定者可能受益于基于系统的教育干预，而不是基于个人的干预。到目前为止，以前的作者提供的模型是对现实世界中涉及信仰极化的更复杂的网络结构的理想化（例如，Madsen和Pilditch， 2018）。因此，本研究的第一个目标是进一步探索ABMs在极化和政治竞选领域的应用，引入更多的参数来考虑更现实的设置。基于Madsen等人（2018）之前的研究结果，因此假设通过将认知和社会偏见引入模型，ec出现得更快、更有规律。第二个目标是澄清在哪种情况下，网络规模和信息密度（Madsen等人，2018）或代理的归纳偏差（Griffiths和Kalish， 2007）与信念极化的形成有关。为了提高生态效度，本调查将涉及真实参与者，并进行迭代学习任务。假设网络规模和归纳偏差都与信念极化的出现呈正相关。

1. OverviewBelief polarization in groups is an important challenge for global political (e.g., Levendusky, 2010) and environmental (Cook and Lewandowsky, 2016) affairs. For example, polarized partisan elites shape public beliefs despite questionable argumentation and lack of sufficient supportive evidence (Druckman, Peterson and Slothuus, 2013). Thus, investigating how polarization relates to the formation of precarious beliefs is key to counteracting developing hazards to democracy (Flynn, Nyhan and Reifler, 2017). Social media sources are frequently used to seek information about political content (e.g., Mitchell et al., 2014) and recent research has investigated the role of social media in the context of belief polarization (e.g., Flaxman, Goel and Rao, 2013). Findings illustrate that repetitive exposure towards selectively chosen information (e.g., via Facebook's algorithmically ranked News Feed) reduces consumption of information discordant with prior beliefs (Bakshy, Messing and Adamic, 2016). Collectively, these findings highlight the importance of understanding the mechanisms underlying belief polarization. Comprehending these mechanisms may help policy makers and system architects mediate the emergence of belief polarization. Moreover, it will allow to learn more about potential benefits and rationality related to the formation of diverging extremist beliefs (Levendusky, 2010; Jern, Chang and Kemp, 2014). 2. Focus and Objective The focus of the current proposal is on recent work using agent-based models (ABMs) and iterated learning (e.g., Griffiths and Kalish, 2007; Kirby, Griffiths and Smith, 2014; Cornish et al., 2017; Madsen, Bailey and Pilditch, 2018).Of particular interest is a recent paper by Madsen et al. (2018). In their agent-based simulations, the authors investigated how network structures can lead to the formation of echo chambers (ECs) even under the assumption that all agents are rational (in the sense of forming beliefs in accordance with Bayes theorem). Surprisingly, it was found that ECs emerged despite the absence of cognitive and social biases. In fact, ECs can emerge through interactions between rational agents under very simple assumptions, and their severity was found to be positively related to network size. This finding has important implications for the reviewed literature concerning belief polarization in groups and its consequences. First, it highlights minimal structural aspects (e.g., interaction between agents and network size) that are sufficient to cause the formation of ECs. Second, it suggests that in order to reduce fixation of extremist beliefs, policy makers may benefit from using system-based educational interventions rather than individual-based interventions. So far, the models provided by previous authors were an idealisation of the more complex network structures involved in belief polarization in a real-world setting (e.g., Madsen and Pilditch, 2018). Therefore, the first objective of the present research is to further explore the application of ABMs in the domain of polarization and political campaigning, introducing additional parameters accounting for more realistic settings. Building on previous findings by Madsen et al. (2018), it is thus hypothesised that ECs emerge faster and more regularly via introducing cognitive and social biases to the model. The second objective is to clarify under which circumstances either network size and information density (Madsen et al., 2018) or the agent's inductive biases (Griffiths and Kalish, 2007) relate to the formation of belief polarization. To increase ecological validity, this investigation will involve real participants following an iterated learning task. It is hypothesised that both network size and inductive biases are positively related to the emergence of belief polarization.