Applying Hybrid Feature Selection Methods for Statistical Modelling of Roadside Particle Concentrations (PM2.5 and PNC)

The task of selecting a predictor variable to include in statistical models is enormous. A model built with fewer predictor variables can be more interpretable and less expensive than the one built with many input variables. In this study, the effects of hybrid feature selection methods (genetic algorithms [GA] and simulated annealing (SA) each combined with random forests [RF]) in improving the efficiency of five variants of multiple linear regression models in the prediction of roadside PM_2.5 and particle number count (PNC) concentrations are investigated. The GA-RF and SA-RF selected 9 and 16 variables, respectively, of the 27 predictor variables in the PM_2.5 training data. Thirteen variables were selected by the GA-RF of the 25 possible variables in the PNC training data, while the SA-RF selected 13 variables.The methods selected variables that are nearly the same especially for predicting PNC, while for the PM_2.5 models the SA-RF selected 16 variables and the GA-RF selected only 10 variables. The hybrid feature selection methods eliminated most of the correlated variables, especially the background pollutants and the traffic variables. Whereas the temporal variables and the meteorological variable have been selected in all the cases considered. The statistical performance of the linear models with the selected variables is similar to those developed using the entire predictor variables. The actual benefit derived from this study is the successful reduction in the number of predictor variables by more than half in most of the cases considered. The reduction in the number of variables will eventually result in the reduction of the operational and computational cost of the models without possibly compromising the predictive performance of the models. Also, the reduction in the number of variables will enhance interpretability.