Stevie Chancellor. Research Statement

Transcription

1 Stevie Chancellor Research Statement Research overwhelmingly shows that online communities can promote healthy behaviors and outcomes for management of disease and disorders [1]. Although most use these communities for positive behavior change, other communities advocate for dangerous health behaviors, like pro-eating disorder (pro-ed), a clandestine online group that promotes eating disorders as lifestyle choices rather than as deadly mental disorders. Pro-ED is an example of what I call deviant mental health behaviors, dangerous and socially stigmatized behaviors related to mental health. These deviant behaviors can have contagion effects by spreading negative emotions to those without a disorder [2], yet also provide therapeutic benefits from participation [3]. To better understanding mental health and make interventions in these circumstances, we must first identify these behaviors and understand their impacts on both individuals and platforms. Pro-ED is an excellent case study that demonstrates the challenges in identifying and understanding deviant mental health behaviors. First, these communities are large, in some cases with tens of millions of posts, and manual techniques for understanding, like qualitative and manual techniques, do not scale. Second, these behaviors are deliberately evasive as they avoid platform moderation and social shame from outsiders. Computational approaches in machine learning have made strides in identifying related deviant behaviors, like spam and abusive content. However, data-driven approaches alone oversimplify the complexities of mental disorders and the unique effects these communities have on people and on platforms. Mental disorders are unique to the individual, yet these communities also have larger platform impacts that resist simplistic approaches to intervention. What is needed to solve this is a solution that blends computational approaches with human-centered insights from stakeholders in these domains. I build human-centered algorithms to analyze deviant health behaviors in online communities. I blend methodological contributions from data-driven fields like machine learning, natural language processing, and data science with collaborations and theoretical insights from fields like psychology and sociology. Using the case study of pro-ed, my research builds rigorous applications of data-driven methods that combine human-in-the-loop techniques to scale to millions of posts on social media that accurately identify and assess these deviant mental health behaviors. My work brings data-driven methods to Human Computer Interaction and Social Computing, a human-centered approach to improve Data Science and Machine Learning predictions, and improves our understanding of moderation, normative behavior, and deviance in socio-technical systems. This is exemplified by a research agenda that has produced 8 firstauthored publications in premier conference venues (CHI, CSCW, FAT*), has been awarded three Best Paper Honorable Mention awards [4, 5, 6], and produced numerous collaborations with clinical and industry stakeholders (Yahoo!; Microsoft Research; Center for Disease Control and Prevention; Columbia University; University of Rochester). Drawing on these human-centered data science, my research agenda will use insights about human behavior and mental health to improve data science and applied AI and answer our toughest questions about deviant behavior online. My future work continues this trajectory in human-centered algorithm design to understand new areas of deviant behaviors, both health and beyond, and encourage rigorous and ethically-sound algorithmic design so that we may be able to design compassionate and effective interventions. My work has two areas of impact that I discuss below: the design of human-centered algorithms to understand deviant mental health behaviors and the insights my work provides to better understand online communities and deviance in socio-technical systems. I then describe my plans for future work in new areas of deviant behaviors (both health and beyond), as well as my goals in designing rigorous and ethical algorithms to understand and predict these behaviors.

2 Designing Human-Centered Algorithms for Deviant Behaviors My research agenda focuses on the pro-eating disorder (pro-ed) community as an extended case study of deviant health behaviors. This work includes bringing data-driven methods to study deviant behavior as well as integrating human insights into algorithm design and implementation. Data-Driven Techniques to Study Deviant Behavior: Pro-ED content is multimodal and combines image, text, and hashtag information to signal its intent - consider the subtleties of a thin fashion model tagged to #fashion compared to #thighgap. Current automated approaches lack this contextual sensitivity to identifying subtle behaviors that convey pro-ed intent. I addressed this gap by deploying deep learning to identify multimodal pro-ed content that violate Tumblr s platform guidelines [7]. I combined convolutional neural networks for image analysis (AlexNet) and word embeddings that then pass to a deep neural network for prediction (performance: 89% accuracy, 65% F1). I tuned our input data, annotation schemes, and error analysis to the needs of moderators on the platform, and demonstrated multiple methods for integrating this system into current moderation practices. Additional research draws on models to make sense of community behaviors in pro-ed communities. For mental disorders, individuals mental illness severity (MIS) varies over time, and this can influence their own predispositions to participate in communities as well as provide key insights for developing strategic interventions. I designed a novel statistical method to combine Latent Dirichlet Allocation (LDA) probability distributions with novice/clinician annotations to infer MIS in over 26 million Instagram posts [4]. We then used these markers of severity to forecast MIS up to seven months in advance. We found, alarmingly, that MIS is on the rise on these platforms (+13%/year), and made recommendations for assessing community well-being. This paper was awarded a Best Paper Honorable Mention at CSCW in 2016 [4]. Integrating Human Insight into Algorithm Design: I believe the massive gains of data-driven techniques to process content must be supplemented with strategic human intervention and expertise currently unavailable to us in machine learning and computational linguistics alone. One persistent challenge in large-scale social research is obtaining ground truth, or quality labels on content to give to prediction systems. Domain experts give reliable and accurate labels, but their efforts are difficult to scale across millions of posts. My work develops novel methods of acquiring and labeling ground truth, drawing on the knowledge of domain experts like psychologists and combining these insights with computational methods to scale human labels [4, 7, 8]. In the same paper about MIS that I discussed before, I collaborated with practicing clinical psychologists to annotate 150 topics with a score of their MIS, and scale these ratings to label 26 million posts [4]. This effectively scaled up the ratings of the domain experts to larger sets of data than what can normally be managed by humans, and also provided a method for tracking MIS changes across the community. Another challenge in understanding deviant behavior is identifying communities who avoid detection. For pro-ed, one method of evasion of platform-enforced content moderation is by changing hashtag spellings to avoid hard bans on words - #thighgap is banned on Instagram, so the community moved to the semantically similar #thyghgapp. I designed an innovative combination of algorithmic snowball sampling of hashtags with human curation to track these lexical changes and comprehensively study the pro-ed community [5]. Starting with known pro-ed tags on Instagram, we iteratively sampled tens of millions of posts to identify potential pro-ed hashtags [5]. These hashtags were then curated by domain experts and organized into 2

3 semantically similar hashtag roots and variants (#thighgap is considered the root of #thyghgapp). This approach allowed me to comprehensively analyze the pro-ed community on Instagram. This theme of incorporating human insight into algorithm design extends to all of my work. I also methods like snowball sampling tags on social media sites like Reddit, Instagram, and Tumblr [5,7,8,9,10] as well as search data [6]. I also use expert annotations to rate data for machine learning classes [7,8], tuning the outputs of models with expert advice and judgment [4,10,11], and conducting error analyses on mistakes made by the classifier [6,7,10]. I also use human sensitivity to interpret the outputs of machine learning classifiers, putting it into a clinical [8] and community context [11]. Insights into Online Communities and Deviance in Socio-technical Systems My research also uses data-driven methods to provide insights into core questions about people in these communities. This includes content moderation and management; and understanding norms of behavior and deviance in socio-technical systems. Content Moderation and Management: Pro-ED and other deviant mental health behaviors are moderated on many social networks, yet the impacts of these moderation strategies are not well-understood. I devised the first quantitative study to investigate the impacts of moderation on pro-ed communities on Instagram. I extracted 713 variant hashtags of 17 root tags, all of which received some kind of content moderation in April 2012 by Instagram [5]. Using computational linguistics to analyze 8 million posts, we found that the pro-ed communities had adopted non-standard lexical variants to circumvent these restrictions. These new communities showed 15-30% more likes/comments on posts and discussed more dangerous and toxic topics compared to the communities on the original hashtags. We concluded that moderation efforts were not successful in reducing pro-ed content on Instagram. This paper was awarded a Best Paper Honorable Mention at CSCW in 2016 [5]. Follow-up work explored the roles of self-censorship on platforms by selfdeleting content [9], and developing robust and powerful classifiers to assist moderators in identifying deviant behavior [7]. Normative Behavior in Online Communities and Deviance: Additionally, my research on the pro-ed community adds to a growing conversation about deviance, bad behaviors and norms in online communities. By leveraging computational linguistic techniques, I was able to show computationally that norms matter in understanding how online health communities promote behavior change [11]. I compared traditional content analysis techniques (TF-IDF, log-likelihood ratios) with those better able to understand the context of words (deep neural network word embeddings). I found that, while the literal content words are similar between a weight loss and pro-ed community on Reddit, the surrounding context of the words are different. I proposed that norms be considered when designing machine learning algorithms as well as informing better design for online communities. Other work touches on the importance of norms in signaling intentions with subtle pro-ed content [7] as well as the complexities of anorexia recovery after participation in pro-ed communities on Tumblr [8]. By using human-centered algorithms to understand online communities, we gain precise and nuanced insights into human behavior, and can begin to develop compassionate and effective behavioral interventions. 3

4 Future Research in Human-Centered Algorithms for Deviant Behavior My growing research agenda will examine the socio-technical systems around deviant behavior in online social platforms. To continue this work, I am excited to investigate two new areas of research: New Areas of Deviant Social Computing Research: Extending beyond the pro-ed community, I am interested in studying other examples of deviant health behavior in online communities. One example are communities that advocate for clinically unverified treatments for opioid addiction, like unregulated drugs and off-label use of known substances. Little research exists on which treatments people use, and whether these treatments are effective for opioid addiction recovery. My current submission to CHI [10] begins this work by identifying the first list of these unverified treatments as well as documenting their controversial effectiveness. To support this work going forward, I will apply for funding from agencies like NIH/NIMH as well as foundations and organizations interested in solving the opioid crisis in the US. In addition to health behaviors, I am also interested in studying other marginalized groups. While at Microsoft Research, I analyzed trends of job seeking behaviors in US search engine data [6]. One of the most insightful outcomes of this work was that my human-centered algorithmic approach let us study the behaviors of job seekers in multi-faceted ways. We found, for example, that the rural job seeker, frequently marginalized by media narratives, is surprisingly similar to their urban counterparts in the kinds of jobs they seek [6]. My techniques for analysis not only allow us to study marginalized communities in nuanced ways, but also pushes the boundaries of how we conceptualize these groups: whose behavior is deviant, and whose is ok? What insights can human-centered algorithms provide to these areas? I am excited to begin exploring these questions and contribute to our understandings of human behavior. Developing Ethical and Rigorous Human-Centered Algorithms: High-profile cases, such as Facebook manipulating the News Feed for experimental purposes and authorizing controversial data access, have brought ethics to greater public scrutiny. How do we conduct ethical research of public data on online platforms? Does the sensitivity of the subject influence what research is allowed to happen? What are new standards for rigor and accuracy in predicting these characteristics? My interdisciplinary expertise in Computer Science and Media Studies positions me to be an expert in negotiating how human computer interaction, data science, and data-driven research can conduct ethical and rigorous research in humancentered algorithm design and artificial intelligence (AI). In this direction, I am currently researching how scientists predicting mental health behaviors from social media data and the tensions in this field. I have discovered that these publications identify divergent and sometimes inconsistent methodological choices, and frequently under-explore ethics and privacy dimensions [12]. In a paper recently accepted at ACM Fairness, Accountability, and Transparency (FAT*), I explore these concerns and ethical challenges, drawing from existing literature, and poses various questions to be resolved as this research gains traction [12]. In the future, I will expand this research into a focus on designing rigorous and ethical solutions to these challenges, and I believe the answer is in human-centered and interdisciplinary collaboration. I anticipate funding for this to come from foundations, NSF, and industry partners who want to be at the forefront of designing human-centered AI initiatives. Through this, I hope to encourage data scientists and quantitative researchers to more critically consider the risks of predictive analysis, participants needs, and how to handle these in responsible, rigorous ways. 4

5 References [1] Gunther Eysenbach, John Powell, Marina Englesakis, Carlos Rizo, and Anita Stern. "Health related virtual communities and electronic support groups: systematic review of the effects of online peer to peer interactions." Bmj 328, no (2004): [2] Jeanne B. Martin. "The development of ideal body image perceptions in the United States." Nutrition Today 45, no. 3 (2010): [3] Tobit Emmens and Andy Phippen. "Evaluating Online Safety Programs." Harvard Berkman Center for Internet and Society.[23 July 2011] (2011). [4] Stevie Chancellor, Zhiyuan (Jerry) Lin, Erica Goodman, Stephanie Zerwas, and Munmun De Choudhury. (2016). Quantifying and Predicting Mental Illness Severity in Online Pro-Eating Disorder Communities. In Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work and Social Computing. CSCW Best Paper Honorable Mention [5] Stevie Chancellor, Jessica Pater, Trustin Clear, Eric Gilbert, and Munmun De Choudhury. (2016). #thyghgapp: Instagram Content Moderation and Lexical Variation in Pro-Eating Disorder Communities. In Proceedings of the 19th ACM Conference on Computer-Supported Cooperative Work and Social Computing. CSCW Best Paper Honorable Mention [6] Stevie Chancellor and Scott Counts. (2018). Measuring Employment Demand Using Internet Search Data. In Proceedings of the 35th Annual ACM Conference on Human Factors in Computing Systems. CHI Best Paper Honorable Mention [7] Stevie Chancellor, Yannis Kalantidis, Jessica Pater, Munmun De Choudhury, and David A Shamma (2017). Multimodal Classification of Moderated Online Pro-Eating Disorder Content. In Proceedings of the 34th Annual ACM Conference on Human Factors in Computing Systems. CHI [8] Stevie Chancellor, Tanushree Mitra, and Munmun De Choudhury. (2016). Recovery Amid Pro-Anorexia: Analysis of Recovery in Social Media. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems. CHI [9] Stevie Chancellor, Zhiyuan (Jerry) Lin, and Munmun De Choudhury. (2016). This Post Will Just Get Taken Down : Characterizing Removed Pro-Eating Disorder Social Media Content. In Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems. CHI [10] Stevie Chancellor, George Nitzburg, Andrea Hu, Francisco Zampieri, and Munmun De Choudhury. Discovering Alternative Treatments for Opioid Use Recovery in Social Media. Under submission at CHI. [11] Stevie Chancellor, Andrea Hu, and Munmun De Choudhury. (2018). Norms Matter: Contrasting Social Support Around Behavior Change in Online Weight Loss Communities. In Proceedings of the 35th Annual ACM Conference on Human Factors in Computing Systems. CHI [12] Stevie Chancellor, Michael Birnbaum, Eric Caine, Vince Silenzio, and Munmun De Choudhury. (2019). Ethical Tensions in Inferring Mental Health States from Social Media: Questions and Calls to Action. Accepted at ACM Fairness, Accountability, and Transparency (FAT*), forthcoming. 5