Efficient and Interpretable Detection of Suicidal Ideation in Social Media via GloVe Embeddings, Recursive Feature Elimination, and a Lightweight Transformer

Abstract

Suicidal ideation expressed on social media offers an opportunity for earlier risk screening; however, practical systems must maximize recall, remain computationally tractable, and provide signals that clinicians can meaningfully interpret. We propose a hybrid pipeline that (i) represents tweets using pretrained 300-dimensional GloVe vectors, (ii) aggregates token embeddings with mean and max pooling, (iii) applies Recursive Feature Elimination (RFE) to select 228 informative semantic dimensions from a 600 dimensional pooled representation, and (iv) trains a compact 6-layer transformer encoder on the selected features for binary classification. Experiments on a combined dataset of 145,230 tweets (26.5% suicidal) demonstrate that the proposed model achieves 94.3% accuracy and an F1-score of 0.944 with 95.1% recall. The architecture is resource-efficient, using 23M parameters and delivering 3.7 imes faster inference than BERT-base. Feature analysis indicates that the most predictive dimensions cluster around clinically relevant themes such as hopelessness, psychological pain, isolation, and perceived burden, offering a more transparent bridge between model behavior and suicide risk theory. These results suggest that feature selection paired with lightweight attention models can yield high-performing screening tools suitable for large-scale monitoring when deployed under rigorous ethical and clinical governance.