Presented at EMNLP 2017, this research explores the potential of social media activity as a scalable screening tool for substance use. By leveraging multi-view unsupervised learning to combine Facebook "Likes" with the text of status updates, the authors developed a system capable of predicting tobacco, alcohol, and illicit drug use with state-of-the-art accuracy.
The Challenge: Data Scarcity and Heterogeneity
Substance Use Disorder (SUD) is a critical public health issue, but traditional screening methods are often expensive and difficult to scale. While social media provides a wealth of behavioral data, building predictive models faces two primary hurdles:
- The Data Scarcity Gap: There is a vast amount of unlabeled social media data but very little "ground truth" data (confirmed self-reports or medical diagnoses).
- The Heterogeneity Gap: Effectively merging different data types—such as discrete "Likes" and unstructured status updates—into a single, cohesive user profile is technically challenging.
Methodology: Multi-View User Embedding (MUE)
The researchers introduced Multi-View User Embedding (MUE), a three-stage process designed to create robust mathematical representations (embeddings) of users:
- Single-View Post Embedding (SPE): User status updates are treated as a single document. The authors found that Distributed Bag of Words (D-DBOW) performed best for this task.
- Single-View Like Embedding (SLE): Similarly, a user's list of "Likes" is treated as a document where each "Like" is a "word." D-DBOW again proved superior, likely because it is insensitive to the arbitrary order in which items are liked.
- Multi-View Learning: To merge these views, the team employed Weighted Generalized Canonical Correlation Analysis (wGCCA) and Deep CCA (DCCA). These algorithms identify a shared space where information from both likes and text are maximally correlated, creating a comprehensive "user fingerprint."
Datasets
The study utilized data from the myPersonality project, a large-scale Facebook research dataset:
- Unsupervised Data: 22 million posts from 153,000 users and 1.8 billion "user-like" pairs from 5 million users.
- Supervised Data: A subset of 13,557 participants who self-reported substance use.
- Evaluation Set: A smaller intersection of 896 users who provided both likes, posts, and SUD labels was used to evaluate the multi-view models.
Linguistic Insights
The study identified significant linguistic markers associated with substance use. For instance, high usage of swear words and topics related to "night life" (e.g., party, bar, drink) showed strong positive correlations with SUD. Conversely, topics related to school or work (e.g., paper, class, job) served as negative predictors.
Key Results and Performance
The multi-view models (specifically wGCCA) consistently outperformed single-view baselines and traditional psycholinguistic models like LIWC. By learning features from millions of unlabeled users before applying SUD labels, the models achieved significant performance gains:
- Tobacco Use: 86% AUC (+17% over unigram baseline)
- Alcohol Use: 81% AUC (+14% over unigram baseline)
- Illicit Drug Use: 84% AUC (+18% over unigram baseline)
These results demonstrate that "Likes" are excellent predictors of identity and preference, while "Posts" capture behavior and mood. Combining them provides a holistic view that neither data source can achieve in isolation.
Limitations and Future Directions
While promising, the study notes several limitations. The evaluation was conducted on a relatively small intersection of users (896) and a demographic skewed toward younger individuals (average age 23). Furthermore, the data reflects Facebook behavior from 2007–2012; modern platforms like TikTok or Instagram may require updated embedding strategies. Future work will focus on developing algorithms that can learn from the union of available data, even when a user has only one type of activity (likes or posts) available.
Summary updated