A Hybrid TF-IDF And Roberta Feature Fusion Framework With Bilstm-Attention For Multi-Platform Depression Detection and Severity Classification With Explainable AI

Abstract

Depression is a major global mental health issue, and the analysis of the social-media text can be used as a scalable and non-invasive means of detecting depressive textual features. The majority of previous work usually relies on a single feature representation, focuses on binary classification, or provides inadequate evaluation on different social-media platforms. In this paper, we propose a novel and explainable hybrid approach that combines TF-IDF features and contextual RoBERTa representation followed by bidirectional LSTM with attention mechanism. This framework is capable of performing binary depression-related text classification as well as four-class severity classification using Twitter and Reddit datasets. We investigate cross-platform transfer and supervised fine-tuning in the target domain, applying SHAP and LIME for explaining the models’ predictions. The results show that our framework achieves 95.55% accuracy and 0.9893 AUC-ROC for binary classification on Reddit, and 96.04% accuracy and 0.9953 AUC-ROC for severity classification. Fine-tuning on 20% of the labeled target-domain data leads to 82.90% accuracy in cross-platform setting.