Quantifying Mental Health from Social Media with Neural User Embeddings
Proceedings of the 2nd Machine Learning for Healthcare Conference, PMLR 68:306-321, 2017.
Mental illnesses adversely affect a significant proportion of the population worldwide. However, the typical methods to estimate and characterize the prevalence of mental health conditions are time-consuming and expensive. Consequently, best-available estimates concerning the prevalence of these conditions are often years out of date. Automated approaches that supplement traditional methods with broad, aggregated information derived from social media provide a potential means of furnishing near real-time estimates at scale. These may in turn provide grist for supporting, evaluating and iteratively improving public health programs and interventions. We propose a novel approach for mental health quantification that leverages user em-beddings induced from social media post histories. Recent work showed that learned user representations capture latent aspects of individuals (e.g., political leanings). This paper investigates whether these representations also correlate with mental health statuses. To this end, we induced embeddings for a set of users known to be affected by depression and post-traumatic stress disorder, and for a set of demographically matched ‘control’ users. We then evaluated the induced user representations with respect to: (i) their ability to capture homophilic relations with respect to mental health statuses; and (ii) their predictive performance in downstream mental health models. Our experimental results demonstrate that learned user embeddings capture relevant signals for mental health quantification.