Multi-Modal Social and Psycho-Linguistic Embedding via Recurrent Neural Networks to Identify Depressed Users in Online Forums

Depression is the most common mental illness in the US, with 6.7% of all adults experiencing a major depressive episode. Unfortunately, depression extends to teens and young users as well and researchers have observed an increasing rate in recent years (from 8.7% in 2005 to 11.3% in 2014 in adolescents and from 8.8 to 9.6% in young adults), especially among girls and women. People themselves are a barrier to fighting this disease as they tend to hide their symptoms and do not receive treatments. However, protected by anonymity, they share their sentiments on the Web, looking for help. In this paper, we address the problem of detecting depressed users in online forums. We analyze user behavior in the ReachOut.com online forum, a platform providing a supportive environment for young people to discuss their everyday issues, including depression. We propose an unsupervised technique based on recurrent neural networks and anomaly detection to detect depressed users. We examine the linguistic style of user posts in combination with network-based features modeling how users connect in the forum. Our results on detecting depressed users show that both psycho-linguistic features derived from user posts and network features are good predictors of users facing depression. Moreover, by combining these two sets of features, we can achieve an F1-measure of 0.64 and perform better than baselines.