Abstract
Bio-inspired metaheuristic optimization offers flexible search mechanisms for high-dimensional predictive problems under operational constraints. In administrative risk prediction settings, class imbalance and feature redundancy challenge conventional learning pipelines. This study evaluates a wrapper-based metaheuristic feature selection framework for post-compliance income declaration prediction using real longitudinal administrative records. The proposed approach integrates swarm-inspired optimization with supervised classifiers under a weighted objective function jointly prioritizing minority-class recall and subset compactness. Robustness is assessed through 31 independent stochastic runs per configuration. The empirical results indicate that performance effects are learner-dependent. For variance-prone classifiers, substantial minority-class recall gains are observed, with recall increasing from 0.284 to 0.849 for k-nearest neighbors and from 0.471 to 0.932 for Random Forest under optimized configurations. For LightGBM, optimized models maintain high recall levels (0.935-0.943 on average) with low dispersion, suggesting representational stabilization and dimensional compression rather than large absolute recall improvements. Optimized subsets retain approximately 16-33 features on average from the original 76-variable space. Within the evaluated experimental protocol, the findings show that metaheuristic-driven wrapper feature selection can reshape predictive representations under class imbalance, enabling simultaneous control of minority-class performance and feature dimensionality. Formal institutional deployment and cross-domain generalization remain subjects for future investigation.