Аннотация
Depression-related language on social media provides measurable signals for population-level mental-health research, yet model selection remains sensitive to evaluation protocol, domain shift, class imbalance, and computational constraints. This study benchmarks CNN, LSTM, and transformer encoders (BERT, RoBERTa, DistilBERT, and MentalBERT) for binary depression-indicative versus control classification on a unified corpus of 19,800 English posts/comments aggregated from three platforms (Reddit, Twitter, and Facebook) under a consistent preprocessing pipeline. We report two complementary evaluation protocols: (1) a fixed-split single-run baseline for a comparable snapshot, and (2) a five-seed repeated-run protocol with statistical testing (effect sizes and multiple-comparison correction) to quantify variability and reduce sensitivity to initialization effects. Under repeated-run reporting, MentalBERT achieves the best overall performance (F1 = 0.918 ± 0.005; AUC = 0.962 ± 0.002), while CNN/LSTM baselines show lower robustness under cross-platform transfer. Cross-domain experiments reveal a consistent performance drop relative to in-domain evaluation, confirming non-trivial platform shift and motivating robustness-aware reporting for deployment-oriented settings. In addition to predictive metrics, we report training time, inference latency, and derived throughput to support practical model selection for use cases such as moderation pipelines and screening/triage dashboards.