Combining Sentiment Lexicons and Content-Based Features for Depression Detection

Numerous studies on mental depression have found that tweets posted by users with major depressive disorder could be utilized for depression detection. The potential of sentiment analysis for detecting depression through an analysis of social media messages has brought increasing attention to this field. In this article, we propose 90 unique features as input to a machine learning classifier framework for detecting depression using social media texts. Derived from a combination of feature extraction approaches using sentiment lexicons and textual contents, these features are able to provide impressive results in terms of depression detection. While the performance of different feature groups varied, the combination of all features resulted in accuracies greater than 96% for all standard single classifiers, and the best accuracy of over 98% with Gradient Boosting, an ensemble classifier.