Neuro-semantic graph fusion for explainable depression risk trajectory mapping using a graph-enhanced RoBERTa...

Early identification of depression through social media analytics is frequently compromised by linguistic ambiguities, contextual complexities, and the opaque nature of conventional deep learning architectures. This study proposes Graph-RoBERTa-CL, a hybrid neuro-semantic framework that explicitly integrates three synergistic components: (i) RoBERTa-based Transformer semantic encoding for deep contextual representation, (ii) Graph Attention Networks (GAT) for relational contextual aggregation across semantically related posts, and (iii) Supervised Contrastive Learning (SCL) for latent space regularization under class imbalance, collectively enabling explainable depression risk trajectory mapping from social media corpora.

The framework was trained and validated on the Suicide Detection Dataset (N = 232,074 posts; 78,541 unique users) using rigorous user-level data splitting to ensure zero user overlap between training and test sets, with cross-dataset consistency assessed on an independent Combined Mental Health Dataset (N = 3,900 posts) withheld entirely from training and model selection. Both datasets are drawn from Reddit communities with community-based labels; broader generalizability across platforms, populations, and clinically diagnosed cohorts requires further validation. User posts are represented as nodes in a multi-relational semantic graph, where a GAT selectively aggregates contextual embeddings to capture ruminative behavioral patterns and temporal trajectories. A supervised contrastive objective simultaneously reshapes the latent space to improve class separability under linguistic ambiguity, and SHAP-based explanations alongside graph attention visualizations provide structural transparency aligned with established psycholinguistic theories of depression.

On the internal user-level test set, Graph-RoBERTa-CL achieved Accuracy 91.2%, Precision 89.7%, Recall 93.1%, F1-Score 0.914, and AUC 0.962, substantially outperforming BERT-base (F1: 0.798, AUC: 0.934) and LSTM-Attention (F1: 0.681, AUC: 0.856). On the external independent validation set, the model achieved F1 0.882 and AUC 0.941, with the modest cross-dataset reduction confirming genuine generalizability rather than dataset-specific overfitting. Ablation studies confirm that each architectural component - semantic encoding, graph attention aggregation, and contrastive optimization - contributes independently and synergistically to overall performance.

The proposed framework establishes a rigorous, interpretable, and ethically grounded mechanism for proactive mental health screening in digital ecosystems, designed to function as a human-in-the-loop clinical screening aid rather than a standalone diagnostic tool.