Multi-strategy enhanced dung beetle optimizer for reliable hyperparameter search in multimodal emotion recognition

Hyperparameter selection in multimodal emotion recognition is challenging because heterogeneous and incomplete modalities can produce noisy validation responses, while configurations selected primarily for peak predictive performance may exhibit poor probability calibration, limited robustness, or unnecessary computational cost. Reliable model selection therefore requires jointly considering discrimination, probability quality, robustness to modality loss, and search efficiency under a limited evaluation budget. To address this problem, this study proposes a multi-strategy enhanced dung beetle optimizer (MSA-DBO) that combines opposition-based initialization, diversity- and remaining-budget-dependent behavioral-role allocation, stagnation-triggered bounded chaotic escape, feasibility repair, and an uncertainty-aware validation objective. The method was evaluated against six hyperparameter-search baselines across four controlled MELD-derived feature domains. MSA-DBO achieved the highest mean weighted F1 of 79.85%, exceeding standard DBO by 4.63% points while showing lower between-run variability. It also obtained the lowest Brier score of 0.146 and negative log-likelihood of 0.425, although Gaussian-process Bayesian optimization achieved a marginally lower expected calibration error. Confidence-based deferral increased selective weighted F1 to 85.31% while retaining 84.8% of predictions. Under controlled modality masking, MSA-DBO achieved the highest weighted F1 under each single-modality omission, whereas BOHB performed slightly better when two modalities were simultaneously unavailable and retained an advantage in search efficiency. Overall, the results indicate that MSA-DBO provides a reliability-oriented hyperparameter-search strategy under the present controlled MELD-derived feature-level conditions. Not applicable.