Proactive depression detection from Facebook text and behavior data

Abstract

This paper proposes a proactive method to detect the clinical depression affected person from post and behavior data from Facebook, called text-based and behavior-based models, respectively. For a text-based model, the words that make up the posts are separated and converted into vectors of terms. A machine learning classification applies the term frequency- inverse document frequency technique to identify important or rare words in the posts. For the behavior-based model, the statistical values of the behavioral data were designed to capture depressive symptoms. The results showed that the behavior-based model was able to detect depressive symptoms better than the text-based model. Regarding performance, a detection model using behaviors yields significantly higher F1 scores than those using words in the post. The k-nearest neighbors (KNN) classifier is the best model with the highest F1 score of 1.0, while the highest F1 score of the behavior-based model is 0.88. An analysis of the predominant features influencing depression signifies that posted messages could detect feelings of self-hatred and suicidal thoughts. At the same time, behavioral manifestations identified depressed people who manifested as restlessness, insomnia, decreased concentration.