Detection of Depression in Arabic Social Media: A Comparison of Traditional and Modern Machine Learning Algorithms

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
The British University in Dubai (BUiD)
This study aims to address the research gap in detecting depression from Arabic tweets using the PHQ-9 scale as a framework. The dataset collected was a set of 200,000 tweets from around 20,000 users. A team of psychologists and assistants used a user-based approach to label users as either depressed or not. The data labelling and annotation process involved a user-based evaluation of the tweets to label users as either depressed or not, based on the two target variables of depressed_binary and depressed_multi. Users with scores between 0 and 6 were categorized as not depressed in the depressed_binary variable, while those with scores above six were classified as depressed. For the depressed_multi variable, users with scores ranging from 0 to 2 were labelled as not depressed, scores from 3 to 6 indicated mild depression, scores from 7 to 9 indicated moderate depression and scores of 10 or above represented high depression. Four machine learning models were employed in this study: HGB (Histogram Gradient Boost), GRU (Gated Recurrent Units), LSTM (Long Short-Term Memory), and SVM (Support Vector Machines). The findings revealed that the older models exhibited strong performance in binary classification, while the new models demonstrated competitive results. Future research should focus on exploring and developing newer deep learning models, such as HGB and GRU models, to enhance the accuracy and performance of depression detection in Arabic tweets. Future studies should also investigate strategies to account for the influence of different Arabic dialects and incorporate Arabic colloquialisms in depression detection models.
social media, Arabic tweets, depression detection