DEPRESSION CLASSIFICATION AND PREDICTION SYSTEM FROM USERS' DATA USING & MACHINE LEARNING TECHNIQUES

ABSTRACT Depression detection from social media texts, such as tweets and comments, is crucial for early intervention, potentiallypreventing severe outcomes like suicide. This study focuses on classifying depression intensity using a labelled Twitterdataset, employing four transformer-based pre-trained language models with fewer than 15 million tunable parameters: ElectraSmall Generator (ESG), Electra Small Discriminator (ESD), XtremeDistil-L6 (XDL), and Albert Base V2 (ABV). The modelswere fine-tuned using hyperparameter optimisation and evaluated on three depression intensity classes: severe, moderate,and mild. Performance metrics were calculated, including accuracy, F1 score, precision, recall, and specificity. A comparativeanalysis with Distil BERT, a larger model with 67 million parameters, revealed ESG as the top-performing model with an F1score of 89%, achieving better contextual understanding and faster training. This research underscores the potential oflightweight models for efficient and accurate depression detection, with recommendations for further optimizing ESG forresource-constrained environments.