Regression based hybrid machine learning model performance evaluation on air quality index prediction in Kolkata

Air pollution, especially elevated particulate matter concentrations, presents a substantial risk to public health and environmental sustainability in urban regions. By employing machine learning and hybrid ensemble models, this study develops a robust frame work for predicting the Air Quality Index (AQI). A multi-step imputation method was used to preprocess the dataset containing metrological variables and air contaminants in order to handle missing values. AQI was selected as the target variable. To assess the possibility of target dependency, two feature configurations were taken in to consideration: a full-featured set and a reduced set that excluded PM2.5 and PM10. Multiple models were used, including Linear Regression, Decision Tree, Random Forest, Gradient Boosting, KNN, MLP and LSTM as well as ensemble methods like Voting and Stacking regressor. Baseline models, namely persistence and SMA were incorporated for comparative analysis. To assess the performance RMSE, MAE, MAPE, RMSLE and R² with a temporal train test split were used. The Voting regressor achieves the lowest RMSE (10.938) and highest R² (0.974), while the Stacking regressor offers the lowest MAE and MAPE demonstrating the superior performance of ensemble models. The LSTM model captures temporal patterns but performs below ensemble models. Models with fewer features perform noticeably worse, underscoring the significance of particulate matter. SHAP analysis shows PM2.5 and PM10 as the most influential features while robustness analysis supports stable performance.