Rhino optimizer for hyperparameter tuning in depression detection

Abstract

Depression detection using machine learning (ML) has gained increasing attention; however, model performance remains highly sensitive to hyperparameter configuration. This study introduces the Rhino optimizer, a novel bio-inspired metaheuristic algorithm for hyperparameter optimization (HPO) inspired by behavioral mechanisms observed in Rhinoceros movement and survival strategies. The optimizer integrates Lévy-flight exploration, elite-guided exploitation, wallowing-based escape behavior, and migration-based repositioning to balance exploration and exploitation while reducing premature convergence. The proposed approach was evaluated on a transformer-based depression-detection model using F1-score as the primary optimization objective due to the class imbalance commonly present in mental-health datasets. Performance was compared against traditional search methods, classical metaheuristics, and modern bio-inspired optimization algorithms under an identical evaluation budget. Rhino optimizer achieved a best observed F1-score of 0.8005, equivalent to the sine-cosine algorithm (SCA) and grey wolf optimization (GWO) with the same evaluation budget. Consistent convergence behavior and statistical analyses suggested that the proposed Rhino optimizer may serve as a competitive approach for HPO in depression detection and that behavior-inspired optimization strategies might be useful in machine-learning applications with limited computational resources.