Predictive Modeling of Urban Air Quality Using Machine Learning

Abstract

Air pollution is a critical environmental concern adversely affecting public health in urban areas worldwide. Accurate prediction of air quality enables timely intervention by environmental agencies and policymakers. This study presents a comprehensive comparative analysis of ensemble machine learning methods for predicting urban air quality using the publicly available UCI Air Quality Dataset, containing 9,358 hourly instances of chemical sensor measurements collected in a heavily polluted Italian city from March 2024 to February 2025. Five algorithms were evaluated: Random Forest (RF), Gradient Boosting Machine (GBM), XGBoost, LightGBM, and CatBoost, against Ridge Regression and SVR baselines. A systematic pipeline encompassing preprocessing, temporal feature engineering, Bayesian hyperparameter optimization via Optuna, and time-series cross-validation was implemented. Results demonstrate XGBoost achieved the best performance with RMSE = 2.14, MAE = 1.58, and R² = 0.9312. SHAP-based feature importance analysis revealed CO(GT) lag features, C6H6(GT), and NOx(GT) as the most influential predictors.

Keywords

• Computer Science
• Air Quality Prediction
• Ensemble Methods
• XGBoost
• SHAP
• UCI Dataset