Depression Detection from Twitter Data Using Two Level Multi-modal Feature Extraction

Abstract

AbstractThe statistics presented by the World Health Organization attributes depression to be a primary cause of concern globally, leading to suicide in majority of the cases if left undetected. Nowadays, Social media is a great point for its users to express their opinions through text, emoticons, photos or videos thus reflecting their sentiments and moods. This has created an opportunity to study social network for understanding the mental state of the users. Studies show that depression generally has an impact on the writing style and corresponding language use. In addition, user persona on social media can also provide us a lot of information about the mental state of the user. The primary aim of our research is to study user’s persona and posts on Twitter and identify the attributes that may indicate depressive symptoms of online users. We used machine learning approaches and natural language processing techniques for training our data and evaluating the efficiency of our proposed method. We proposed a two-level depression detection in which the social media features, personality trait and sentiment analysis of user’ biography provide us an opportunity to identify suspected depressed users. We combined these attributes with other Linguistic features (N-Gram+TF-IDF) and LDA and achieved an accuracy of 89% using Support Vector Machine classifier. According to our research, proper feature selection and their combinations help in achieving better improvement in performance.KeywordsDepressionNatural language processingMachine learningUser profile analysisTwitterSocial networks