Deep learning models have shown promising results for depression detection using social media data (i.e., Twitter), but the difficulties of maintaining explainability and few-shot adaptation of models for new problems remain an open challenge. Another challenging aspect of the problem of depression detection in social media is the fact that the number of instances belonging to the depressed class are in a minority when compared to the number of instances belonging to the non-depressed class. This, especially, makes it harder for supervised machine learning algorithms to learn and predict depressed class instances.In this study, we proposed a simple solution to this problem by generating \extit{differential embeddings} using the Sentence BERT transformer architecture. More specifically, we proposed a few-shot model that can leverage state-of-the-art (SOTA) representation learning techniques and used it in supervised and unsupervised tasks.We constructed a small set of dysfunctional thought patterns in the embedding space, i.e., a set of clinically-backed depression symptoms. We then used SBERT embedding vectors to measure the similarities between different tweets and anchor points as a distance in the vector space, or fed them directly into the machine learning model. We assessed the capability of our approach on two different datasets. We trained supervised and unsupervised models using different approaches that were derived from Sentence-BERT and the anchor points. Results show that the proposed solution improved SBERT in both supervised and unsupervised tasks.
Read
