SEEDS: Curriculum-Driven Adaptive Regression Contrastive Learning for Robust Evidence-Based Depression Severity Classification

Depression detection from social media faces critical challenges: sparse signals obscured by noisy content, severe class imbalance biasing models toward non-depressive text, and arbitrary feature extraction in end-to-end neural models causing overfitting on small datasets. We propose SEEDS–Sequential Evidence-based Encoder for Depression Severity classification–which integrates evidence-based encoding with curriculum-driven adaptive contrastive learning. SEEDS extracts structured clinical features aligned with BDI-II symptoms and emotional states, replacing arbitrary extraction with interpretable evidence. A multi-task curriculum architecture progressively filters relevant content, providing balanced training data for severity estimation. We introduce adaptive regression contrastive learning to extend contrastive optimization to continuous prediction while dynamically balancing loss convergence. Evaluation on CLEF eRisk, Deptweet, and DepSeverity datasets demonstrates state-of-the-art macro F1 performance, validating SEEDS’s effectiveness across user-level and post-level depression detection. Source code is available at https://github.com/candleMind/SEEDS .