A two stage feature selection method using AdjcorT and radial basis function neural network for enhance classification of daily PM2.5 concentration

This study addresses the critical challenge of multicollinearity in air quality datasets by proposing a two-stage feature selection method, AdjcorT, integrated with the Radial Basis Function Neural Network (RBFNN) to improve the classification of daily PM2.5 concentrations. Multicollinearity, characterized by strong correlations among predictor variables, can distort model performance and obscure the true importance of features, resulting in less reliable classifications. The AdjcorT method explicitly accounts for both positive and negative correlations, effectively mitigating the adverse effects of multicollinearity during feature selection. The AdjcorT-RBFNN model was evaluated using datasets from Shah Alam and Banting. In Shah Alam, AdjcorT-RBFNN outperformed other models, including RBFNN, Lasso-RBFNN, mRMR-RBFNN, and ReliefF-RBFNN, while in Banting, Lasso-RBFNN showed better predictive performance. However, AdjcorT-RBFNN demonstrated competitive accuracy with significantly faster computational time. The study highlights the critical role of feature selection that considers high correlations among features to improve predictive accuracy. Key predictors for next-day PM2.5 in Shah Alam include NO2, PM2.5, PM10, CO, O3, wind speed, SO2, and wind direction. In Banting, significant predictors include humidity, PM2.5, wind direction, temperature, PM10, NO2, wind speed, and CO. The research underscores the importance of addressing multicollinearity, as both AdjcorT and Lasso consistently identified core predictors such as PM2.5, PM10, and NO2. Although Lasso-RBFNN achieved higher accuracy in Banting, AdjcorT-RBFNN was more computationally efficient while delivering similarly robust results. Furthermore, simulations demonstrate AdjcorT-RBFNN’s effectiveness in managing highly correlated variables, especially under strong correlation conditions. This study contributes to the development of more accurate and computationally efficient models for air quality classification, emphasizing the need to select appropriate feature selection methods based on dataset characteristics, particularly when multicollinearity is present.