A Comparison of Machine Learning Depression Detection Models on Diverse Datasets

Depression identification via social media monitoring is an important step toward early intervention, providing a means to detect those at risk through their online behavior. This study uses machine learning methods to label depression posts on Reddit and Twitter. The datasets were thoroughly preprocessed, with text cleaning, removal of stopwords, and TF-IDF vectorization, providing high-quality input features. Three models- Naive Bayes, Logistic Regression, along with Random Forest which were trained and tested. Among them, Random Forest showed the best accuracy with 94.96% on the Reddit test set and 68.80% on the Twitter test set. These outcomes indicate the difficulty of cross-platform generalization owing to varying writing styles, lengths of posts, and user participation. Furthermore, a real-time prediction function was implemented to classify user-input text in real time. The research confirms that although machine learning models efficiently identify depressive content, the performance is heavily reliant on dataset properties. Future work will investigate more sophisticated deep learning strategies like BERT to increase the accuracy of classification and ensure platform robustness. In addition, continuous learning mechanisms will be incorporated to enable the model to retrain on fresh data and learn over time. This contribution is part of the expanding body of research on AI-based mental health evaluation, with potential uses in psychological screening and early intervention systems.