Multi-View Unsupervised User Feature Embedding for Social Media-based Substance Use Prediction

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2275–2284 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Multi-View Unsupervised User Feature Embedding for Social Media-based Substance Use Prediction Tao Ding Department of Information Systems University of Maryland, Baltimore County taoding01@umbc.edu Warren K. Bickel Addiction Recovery Research Center Virginia Tech Carilion Research Institute wkbickel@vtc.vt.edu Shimei Pan Department of Information Systems University of Maryland, Baltimore County shimei@umbc.edu Abstract In this paper, we demonstrate how the state-of-the-art machine learning and text mining techniques can be used to build ef- fective social media-based substance use detection systems. Since a substance use ground truth is difficult to obtain on a large scale, to maximize system performance, we explore different unsupervised feature learning methods to take advantage of a large amount of unsupervised social me- dia data. We also demonstrate the benefit of using multi-view unsupervised feature learning to combine heterogeneous user information such as Facebook “likes” and “status updates” to enhance system per- formance. Based on our evaluation, our best models achieved 86% AUC for pre- dicting tobacco use, 81% for alcohol use and 84% for illicit drug use, all of which significantly outperformed existing meth- ods. Our investigation has also uncovered interesting relations between a user’s so- cial media behavior (e.g., word usage) and substance use. 1 Introduction A substance use disorder (SUD) is defined as a condition in which recurrent use of substances such as alcohol, drugs and tobacco causes clin- ically and functionally significant impairment in an individual’s daily life (SAMHSA, 2015). Ac- cording to the 2014 National Survey on Drug Use and Health, 1 in 10 Americans age 12 and older had a substance use disorder. Substance use also costs Americans more than $700 billion a year in increased health care costs, crimes and lost pro- ductivity (NIDA, 2015). These days, people also spend a significant amount of time on social media such as Twitter, Facebook and Instagram to interact with friends and families, exchange ideas and thoughts, pro- vide status updates and organize events and activ- ities. The ubiquity and widespread use of social media underlines the needs to explore its intersec- tion with substance use and its potential as a scal- able and cost-effective solution for screening and preventing substance misuse and abuse. In this research, we employ the state-of-the-art machine learning and text mining algorithms to build automated substance use prediction systems, which can be used to identify people who are at risk of SUD. Since SUD data are often expen- sive to obtain at a large scale, to maximize system performance, we focus on methods that employ unsupervised feature learning to take advantage of a large amount of unsupervised social media data. Previous research in Machine Learning, Im- age Processing, Speech and Natural language Pro- cessing has shown that to be able to utilize a large amount of unsupervised data is one of the most re- liable ways to achieve good performance (Le et al., 2011; Lee et al., 2009; …

Content truncated to 500 words