Preventing Catastrophic Forgetting in New Tasks: Development of Artificial Intelligence Based Models for Depression Detection in Continual Learning Settings

The surge in social media users has made it a significant platform where people’s behaviors are reflected in their posts and comments. Individuals exhibiting signs of depression may unintentionally express negative or distressing texts seeking support from others. Consequently, analyzing social media texts enables the detection of potential signs of depression among users and provides an opportunity to offer assistance and support to those in need. This article presents a comprehensive study on detecting depression-related texts in social media comments using deep learning models, namely XLNET, DistilBERT, and ALBERT, alongside traditional machine learning models such as Naive Bayes, Support Vector Machine, and Random Forest. The research aims to compare the performance of these models in identifying linguistic patterns indicative of depression within user-generated content on social media platforms. The study utilizes a dataset of social media comments collected from Reddit containing both depressive and non-depressive content, carefully annotated for training and evaluation purposes. Furthermore, text summarization enhances the efficiency of the proposed models by condensing lengthy text into concise representations. Of all the proposed models, XLNET with summarization outperforms other models and gives 63.33% accuracy. In addition, we test the task-incremental learning capability of the developed transformer models in a continual learning setting on a new task using Elastic Weight Consolidation which prevents catastrophic forgetting. The results demonstrate that the models perform better in continual learning settings by preserving important weights and balancing old and new knowledge.