Bioinspired Optimization for Feature Selection in Post-Compliance Risk Prediction

Area of Science:

• Computational intelligence and machine learning
• Data science and predictive analytics
• Administrative science and public policy

Background:

• Class imbalance and feature redundancy pose challenges in administrative risk prediction.
• Conventional learning pipelines struggle with high-dimensional data and operational constraints.
• Bio-inspired metaheuristic optimization offers flexible search mechanisms for complex predictive tasks.

Purpose of the Study:

• To evaluate a wrapper-based metaheuristic feature selection framework for post-compliance income declaration prediction.
• To integrate swarm-inspired optimization with supervised classifiers using a weighted objective function.
• To jointly prioritize minority-class recall and subset compactness in predictive models.

Main Methods:

• Utilized real longitudinal administrative records for income declaration prediction.
• Implemented a wrapper-based metaheuristic feature selection framework.
• Integrated swarm-inspired optimization with supervised classifiers (k-nearest neighbors, Random Forest, LightGBM).
• Assessed robustness through 31 independent stochastic runs per configuration.

Main Results:

• Metaheuristic feature selection significantly improved minority-class recall for variance-prone classifiers (e.g., k-nearest neighbors, Random Forest).
• Optimized models for LightGBM maintained high recall with reduced feature dimensionality (16-33 features from 76).
• Performance gains were learner-dependent, indicating the importance of classifier choice.
• The approach demonstrated simultaneous control over minority-class performance and feature dimensionality.

Conclusions:

• Metaheuristic-driven wrapper feature selection effectively reshapes predictive representations for imbalanced datasets.
• The framework enables simultaneous optimization of minority-class performance and feature dimensionality.
• Findings suggest potential for improved administrative risk prediction models.
• Further investigation into institutional deployment and cross-domain generalization is warranted.