Improving Suicide Ideation Detection Through Feature Engineering and Machine Learning

Suicidal ideation detection from social media text is a growing area of research, yet many existing approaches suffer from limited contextual understanding, low interpretability, and inadequate scalability. This study addresses these challenges by proposing a comprehensive machine learning framework that balances performance, interpretability, and real-world applicability. Using a curated Reddit dataset (comparing “SuicideWatch” and “nonsuicide” posts), we evaluate both traditional feature extraction methods (TF-IDF, Bag-of-Words, and N-Grams) and contextual embeddings generated by BERT. These features are tested across five established classifiers: Random Forest, SVM, Logistic Regression, XGBoost, and LightGBM. Our results show that BERT combined with Logistic Regression achieves the highest F1-score (0.9392), outperforming both traditional models and ensemble methods. Additionally, we assess model robustness under varying dataset sizes and address class imbalance through targeted preprocessing and evaluation strategies. While our framework demonstrates strong performance, it is currently limited to English-language Reddit data and requires validation across platforms and demographics. This work contributes a reproducible and scalable pipeline to support mental health monitoring in online environments.