DDdeep: deep learning-based text analysis for depression illness detection on social media posts

Abstract

Abstract Recently, depression has been raised as one of the most popular mental health disorders in the world. Also, social networks can be considered a valuable resource for mental health research due to the tendency of users for sharing their thoughts and feelings. On the other hand, text analysis of user posts relying on neural networks for such research is increasing. Neural networks have recently achieved significant success in text analysis because of the ability to automatically extract distinguishing features from data. However, neural networks are ignored the temporal and sequential nature of users' posts on social networks which affects the accuracy of the results. This shortcoming prompted us to present a more efficient method considering the sequential and temporal nature of social media users' posts. Thus, we have proposed a deep learning-based hybrid method called DDdeep to handle the mentioned challenge. There are three main features in our method, which are (1) text analysis relying on the temporal and sequential nature of posts, (2) identifying depressed users only by considering how users use language, and (3) remembering decisions because of the dependence of each post on previous posts. The DDdeep method has integrated a convolutional neural network (CNN) to extract more important features and long-short term memory (LSTM) to remember previous decisions. Our method identifies the depressed users by 78% precision, 70% recall, and 73% F1-score. Therefore, the evaluation results of our method are acceptable and competitive compared to other valid methods in this field.