Enhancing the Prediction of Multiple Ozone Metrics Using Genetic Algorithm-Based Feature Selection for the Multi-Target Regression of the Environmental AQ-Bench Dataset

Predicting multiple air quality metrics from high-dimensional environmental datasets is a significant challenge hampered by the "curse of dimensionality". The present study introduces and evaluates three novel Genetic Algorithm (GA)-based feature selection methodologies designed specifically for Multi-Target Regression (MTR) tasks using the AQ-Bench dataset. The proposed wrapper-based approach integrates GAs with MTR models to identify optimal feature subsets. The results demonstrate a substantial reduction in feature dimensionality, by up to 61.6%, while concurrently improving predictive performance over baseline models. This research establishes a practical framework for practitioners, showing that a common feature subset (GA-FS-MTR) is effective for correlated targets, whereas a per-target approach (GA-FS-TARGET) excels when precision for heterogeneous targets is required. A key finding of the present study is the identification of a structural sensitivity in complex models like the Ensemble of Regressor Chains (ERC), where global optimization can inadvertently remove features vital for its chained architecture. This work validates GA-based feature selection as an effective tool for optimizing MTR models in environmental science and provides a strategic guide for its implementation.