An explainable ensemble model for revealing the level of depression in social media by considering personality traits and sentiment polarity pattern

Early detection of depression in mental health is crucial for better intervention. Social media has been extensively used to examine users’ behavior, motivating researchers to develop an automatic depression detection model. However, the accuracy and clarity of the reasons behind the detection results still need to be improved. Current research focuses primarily on syntactic and semantic information in user-posted texts, while other aspects of users’ psychological characteristics are often overlooked. Therefore, this study addresses the gap by proposing a novel model integrating personality traits and sentiment polarity patterns into an explainable ensemble model. Specifically, we developed two base learners for the averaged and meta-ensemble learning strategy. The first learner employed the Robustly Optimized BERT Pre-training Approach (RoBERTa). For the second learner, we combined the Random Forest and Bidirectional Long Short-Term Memory (RF-BiLSTM) methods to effectively handle the combination of personality traits and sequential information in sentiment polarity patterns. These additional features are obtained by performing domain adaptation for personality prediction and sentiment analysis using a lexicon-based model. Based on the experimental results, our ensemble model improved depression detection results by leveraging the strengths of each base learner. Our model advanced the state-of-the-art, outperforming existing models with an increase in accuracy and F1-score of 4.14% and 2.99%, respectively. The model successfully enhanced the interpretability of detection results, providing a more comprehensive understanding of the factors underlying depressive symptoms. This research highlights the potential of considering alternative additional features as a promising avenue for enhancing depression detection in social media.