Auditing Social Media Algorithmic Pathways to Measure Prevalence of Online Misinformation Related to Opioid Misuse

摘要：阿片类药物滥用已成为美国的一种公共卫生流行病，超过70%的患者服用阿片类药物。 患有阿片类药物使用障碍(OUD)的患者从未接受过任何治疗。更少的人接受药物治疗 对于成瘾治疗(MAT)-治疗的黄金标准和安全、经济有效的方法来降低 服药过量，同时提高持续康复的可能性。由于围绕阿片类药物滥用的耻辱，在- 个人经常寻求非传统的恢复方式，例如使用在线资源，特别是社交媒体， 尤其是像Twitter这样的微博网站。然而，社交媒体平台上经常充斥着垫子失踪的情况- 形成(MATM)，对复苏构成严重障碍。此外，网上虚假信息的有害影响 在社交网站上推动内容管理或推荐的算法设计进一步加剧了 媒体网站。然而，关于理解导致健康错误信息的算法路径的研究很少，而且这种研究还不多见。 最近的阿片类药物滥用几乎是不存在的。此R21提案将通过进行形成性 通过使用稳健的审计方法和经过严格验证的机器学习(ML)进行研究 技术，揭示了在我们的药物和治疗领域中的一个未被探索的现象-算法 在在线社交媒体系统中策划MATM，特别是Twitter-使用最广泛的社交媒体之一 共享和寻找网络信息的平台。这项工作通过利用 团队在解决推动该提案的两个关键技术挑战方面的开创性研究：a)构建 审核用于精选、推荐或过滤信息的黑盒平台算法的计算方法 2)开发ML技术，检测预先存在的或紧急出现的在线错误信息- 提顿。借鉴算法审计工作的进展和PI自己成功的审计研究设计，目标1将 构建工具和方法来审核Twitter上各种不同的MATM搜索和推荐算法 OU个人用户特征和算法输入。所开发的方法将足够通用 在其他社交媒体平台上都能适应。在目标2中，我们将利用这些方法进行 在Twitter上进行了一系列精心控制的详尽审计实验，以调查其搜索和推荐- TION算法向表面MATM的趋势。我们还将开发和评估可以自动执行的ML方法 确定收集的社交媒体帖子是否包含MATM。最后，在目标3中，我们将开发一种混合- 与Twitter上的参与者一起定量和定性验证我们的审计结果的方法 滥用阿片类药物。该项目汇集了一个由计算机科学家和临床心理学家组成的多学科团队-- 在社交媒体分析和招聘、在线算法审计、物质使用障碍、 机器学习和自然语言处理。我们创造的知识将为未来的重新定位奠定基础 在早期发现危险的OUD行为方面进行搜索，了解在线信息环境在 加剧或预防这些风险，并启动循证干预措施以减轻此类风险。

Abstract: Opioid misuse has become a public health epidemic in the United States with more than 70% of indi- viduals with an opioid use disorder (OUD) never receiving any sort of treatment. Even fewer receive medications for addiction treatment (MAT)—the gold standard for treatment and a safe, cost-effective way to reduce the risk of overdose while improving the likelihood of sustained recovery. Due to the stigma surrounding opioid misuse, in- dividuals often seek non-conventional ways to recover, such as using online resources, specifically social media, and in particular microblogging sites like Twitter. However, social media platforms are often rife with MAT misin- formation (MATM), posing a serious barrier to recovery. Moreover, the harmful effects of online misinformation are further exacerbated by the design of the algorithms that drive content curation or recommendation on social media sites. Yet, research on understanding algorithmic pathways to health-misinformation is rare and that re- lated to opioid misuse is practically non-existent. This R21 proposal will address this gap by conducting formative research through the use of robust audit methodologies coupled with rigorously validated machine learning (ML) techniques, to lay bare an unexplored phenomena in the OUD medication and treatment domain—algorithmically curated MATM in online social media systems, specifically Twitter—one of the most widely used social media platforms for sharing and seeking OUD information. The work advances this research agenda by leveraging the team’s pioneering research in addressing two of the key technical challenges driving this proposal: a) building computational approaches to audit black-box platform algorithms that curate, recommend, or filter information viewed by end users; and 2) developing ML techniques that detect pre-existing or emergent online misinforma- tion. Drawing from advances in algorithmic audit work and PI’s own successful audit study designs, Aim 1 will build tools and methodologies to audit search and recommendation algorithms for MATM on Twitter across vari- ous individual user characteristics and algorithmic inputs. The developed methodologies will be generic enough to be adaptable across other social media platforms. In Aim 2, we will leverage these methodologies to conduct an exhaustive set of carefully controlled audit experiments on Twitter to investigate it’s search and recommenda- tion algorithms’ tendency to surface MATM. We will also develop and evaluate ML methods that can automatically determine whether the collected social media posts contain MATM. Finally, in Aim 3 we will develop a mixed- methods approach to quantitatively and qualitatively validate our audit results with participants on Twitter who misuse opioids. The project brings together a multidisciplinary team of computer scientists and a clinical psychol- ogist, with expertise in social media analytics and recruitment, online algorithmic audits, substance use disorders, machine learning, and natural language processing. The knowledge we produce will set the stage for future re- search in early detection of risky OUD behaviors, understanding the role of the online information environment in exacerbating or preventing OUD risks and launching evidence-based interventions to mitigate such risks.