Abstract
Recent studies indicate that open and closed vocabulary natural language processing (NLP) approaches can detect individuals dealing with depression, by studying the language they use. This has motivated researchers to build models capable of detecting depression via social media. However, given limited user data, adopting the existing approaches will face instability during training, and ultimately, will result in poor performance of the learned model. To address the aforementioned issue, we take advantage of the content of users' social circles to detect depression in individuals. Accordingly, we begin by constructing the first social media depression dataset on Twitter, containing tweets of users and their online social group members (friends). Subsequently, we introduce MentalSpot, a deep learning model aimed at detecting depression in users with limited data. In MentalSpot, a triplet training architecture is employed to create novel embeddings that reflect the existence of depression in users. Furthermore, users' content is joined with their most similar friends' and then fed to a convolutional neural network model for depression detection. Extensive experiments on users with limited data showed that MentalSpot yielded superior performance compared to the state-of-the-art methods that solely took into account user content to detect depression in social media users.