Coordinated inauthentic behavior is used as a tool on social media to shape public opinion by elevating or suppressing topics using systematic engagements -- e.g. through *likes* or similar reactions. In an honest world, reactions may be informative to users when selecting on what to spend their attention: through the wisdom of crowds, summed reactions may help identifying relevant and high-quality content. This is nullified by coordinated inauthentic liking. To restore wisdom-of-crowds effects, it is therefore desirable to separate the inauthentic agents from the wise crowd, and use only the latter as a voting *jury* on the relevance of a post. To this end, we design two *jury selection procedures* (JSPs) that discard agents classified as inauthentic. Using machine learning techniques, both cluster on binary vote data -- one using a Gaussian Mixture Model (GMM JSP), one the k-means algorithm (KM JSP) -- and label agents by logistic regression. We evaluate the jury selection procedures with an agent-based model, and show that the GMM JSP detects more inauthentic agents, but both JSPs select juries with vastly increased correctness of vote by majority. This proof of concept provides an argument for the release of reactions data from social media platforms through a direct use-case in the fight against online misinformation.

翻译：协同造假行为被用作社交媒体上的工具，通过系统性互动（例如“点赞”或类似反应）来提升或压制话题，从而塑造公众舆论。在一个诚实的环境中，反应可能对用户选择关注内容具有参考价值：通过群体智慧，累计的反应有助于识别相关且高质量的内容。然而，协同造假点赞会抵消这一效果。为恢复群体智慧效应，有必要将造假参与者与诚实群体区分开来，并仅以后者作为判断帖子相关性的投票“评审团”。为此，我们设计了两种评审团选择程序（JSPs），用于剔除被分类为造假参与者的代理。利用机器学习技术，这两种程序对二值投票数据进行聚类——一种使用高斯混合模型（GMM JSP），另一种使用k-means算法（KM JSP）——并通过逻辑回归对代理进行标记。我们通过基于代理的模型评估了这些评审团选择程序，结果表明，GMM JSP检测到的造假代理更多，但两种JSP所选评审团的多数投票正确性均显著提高。这一概念验证通过一个直接的应用场景，为社交媒体平台发布反应数据提供了论据，以支持对抗在线虚假信息的行动。