Learning Language and Multimodal Privacy-Preserving Markers of Mood from Mobile Data

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4170–4187 August 1–6, 2021. ©2021 Association for Computational Linguistics 4170 Learning Language and Multimodal Privacy-Preserving Markers of Mood from Mobile Data Paul Pu Liang1⋆, Terrance Liu1⋆, Anna Cai1, Michal Muszynski1, Ryo Ishii1, Nicholas Allen2, Randy Auerbach3, David Brent4, Ruslan Salakhutdinov1, Louis-Philippe Morency1 1Carnegie Mellon University 2University of Oregon 3Columbia University 4University of Pittsburgh {pliang,terrancl,annacai,mmuszyns,rishii,rsalakhu,morency}@cs.cmu.edu nallen3@uoregon.edu rpa2009@cumc.columbia.edu brentda@upmc.edu Abstract Mental health conditions remain underdiag- nosed even in countries with common access to advanced medical care. The ability to accu- rately and efficiently predict mood from eas- ily collectible data has several important im- plications for the early detection, intervention, and treatment of mental health disorders. One promising data source to help monitor human behavior is daily smartphone usage. However, care must be taken to summarize behaviors without identifying the user through personal (e.g., personally identifiable information) or protected (e.g., race, gender) attributes. In this paper, we study behavioral markers of daily mood using a recent dataset of mobile behav- iors from adolescent populations at high risk of suicidal behaviors. Using computational models, we find that language and multimodal representations of mobile typed text (spanning typed characters, words, keystroke timings, and app usage) are predictive of daily mood. However, we find that models trained to pre- dict mood often also capture private user iden- tities in their intermediate representations. To tackle this problem, we evaluate approaches that obfuscate user identity while remaining predictive. By combining multimodal repre- sentations with privacy-preserving learning, we are able to push forward the performance- privacy frontier. 1 Introduction Mental illnesses can have a damaging permanent impact on communities, societies, and economies all over the world (World Health Organization, 2003). Individuals often do not realize they are at risk of mental disorders even when they have symptoms. As a result, many are late in seeking professional help and treatment (Thornicroft et al., 2016), particularly among adolescents where sui- cide is the second leading cause of death (Curtin ⋆first two authors contributed equally. Real-time assessment Decentralized multimodal mobile device data Aggregate Privacy-preserving representation learning Figure 1: Intensive monitoring of behaviors via adoles- cents’ natural use of smartphones may help identify real-time predictors of mood in high-risk youth as a proxy for suicide risk. While smartphones provide a valuable data source spanning text, keystrokes, app us- age, and geolocation, one must take care to summarize behaviors without revealing user identities through per- sonal (e.g., personally identifiable information) or pro- tected attributes (e.g., race, gender) to potentially adver- sarial third parties. and Heron, 2019). In addition to deaths, 16% of high school students report having serious suicidal thoughts each year, and 8% of them make one or more suicide attempts (CDC, 2015). This problem is particularly exacerbated as an “echo pandemic” of mental health problems have arisen in the wake of the COVID-19 pandemic (Inkster et al., 2021; Saha et al., 2020). Intensive monitoring of behaviors via adolescents’ …

Content truncated to 500 words