Early depression detection in social media based on deep learning and underlying emotions

Depression is a challenge to public health, frequently related to disability and one of the reasons that lead to suicide. Many of the ones who suffer depression use social media to obtain information or even to talk about their problem. Some studies have proposed to detect potentially depressive users in these online environments. However, unsatisfactory effectiveness is still a barrier to practical application. Hence, we propose a method of early detection of depression in social media based on a convolutional neural network in combination with context-independent word embeddings and Early and Late Fusion approaches. These approaches are experimentally evaluated, considering the importance of the underlying emotions encoded in the emoticons. The results show that the proposed method was able to detect potentially depressive users, reaching a precision of 0.76 with equivalent or superior effectiveness in relation to many baselines ($F_1(0.71)$). In addition, the semantic mapping of emoticons allowed to obtain significantly better results, including higher recall and precision with gains of 46.3% and 32.1%, respectively. Regarding the baseline word embedding approach, the higher recall and precision gains of our semantic mapping of emoticons were 14.5% and 40.8%. In terms of overall effectiveness, this work advanced the state-of-the-art, considering both individual embeddings and the fusion-based approaches. Moreover, it is demonstrated that emotions expressed by depressed people and encoded through emoticons are important suggestive evidence of the problem and a valuable asset for early detection.