Social-media platforms provide abundant signals related to mood disorders, yet building reliable supervised models is hindered by limited expert annotations and heterogeneous, noisy language. This paper introduces a two-stage framework for mood-state classification (mania, depression, normal) that leverages large-scale unlabeled posts while preserving evaluation rigor on a strictly held-out clinician-labeled benchmark ( \(G^{500}_{\textrm{test}}\) ). In Stage 1, we generate pseudo-labels using a Flan-T5 self-consistency scheme that samples multiple label proposals per post and aggregates them by majority vote to retain high-agreement instances. This yields markedly cleaner supervision, reaching 0.870 accuracy and 0.863 macro-F1 on \(G^{500}_{\textrm{test}}\) , improving over the strongest labeling baselines (0.538 accuracy and 0.446 macro-F1) by +0.332 and +0.417 absolute points (+61.7% and +93.5%, respectively). Importantly, worst-class robustness (Min-F1) increases from 0.165 to 0.830 (+0.665 absolute; 5.03 \(\varvec{\times }\) , i.e., +403%), clarifying that the large relative gain is driven by a low baseline Min-F1. In Stage 2, we cast model selection as a multi-objective optimization problem that jointly maximizes macro-F1 and worst-class F1 while minimizing inference latency, and solve it using Bayesian optimization with qEHVI (via BoTorch). The optimized configurations yield +4.9% macro-F1 and +7.3% minimum F1 with a 33% latency reduction relative to an untuned baseline (0.803 macro-F1, 0.772 Min-F1, latency 138.6), providing a practical accuracy–efficiency trade-off. To quantify uncertainty and confirm that observed improvements are statistically supported, we perform paired significance analyses on \(G^{500}_{\textrm{test}}\) and report 95% bootstrap confidence intervals. Extensive experiments reveal Pareto-optimal solutions that are appropriate for deployment under resource constraints and demonstrate steady improvements across evaluation metrics.
