Automatic classification of depressive users on Twitter including temporal analysis

AbstractIn recent years, identifying traits of mental illness on social media platforms have caught researchers ’ attention. Unfortunately, different mental illnesses have similar symptoms, which makes detection very challenging. The present study conducts an examination of depressed users on the social media platform Twitter, utilizing the CLPsych 2015 database. In addition to conducting an analysis of sym ptoms and emotions as in current research, we have incorporated temporal analysis. This innovative approach has enabled us to differentiate between two prevalent forms of depression: Major Depressive Disorder and Persistent Depressive Disorder. Features proposed for each disorder were used to improv e the results. A Convolutional Neural Network and an XGBoost model were used to classify users into depressive and non-depressive users using GloVe and TF-IDF as word embeddings, respectively. We obtained better results 0.8843 accuracy, 0.8751 F1-score, 0.885 precision, 0.8704 recall, and 0.862 AUC for the CNN model. The results were higher than the state of the art except for the AUC metric when the features of each disease were added. Both algorithms had similar behaviors, but XGBoost had a lower variance. Our study shows that incorporating temporal analysis is an important feature in diagno sing depression on social media because depressed users may experience different symptoms based on the type of depression they experience.
Source: Network Modeling Analysis in Health Informatics and Bioinformatics - Category: Bioinformatics Source Type: research