A comparison of ranking filter methods applied to the estimation of NO2 concentrations in the Bay of Algeciras (Spain)

This study presents a comparison between sixteen filter ranking methods applied to a real air pollution problem. Adaptations of the Minimum-Redundancy-Maximum-Relevance (mRMR) algorithm to use the Spearman's rank correlation, the kernel canonical correlation analysis, the Hilbert-Schmidt independence criterion, correntropy, the Pearson's correlation and the distance correlation are included among them. These methods were compared by estimating the hourly NO2 concentrations at three monitoring stations located in the Bay of Algeciras (Spain). The estimation models were generated using Bayesian regularized artificial neural networks. Different estimation cases were tested for each ranking method. Finally, results were statistically compared to determine which filter ranking strategy produced the best performing model in each case. The proposed estimation scenarios showed how mRMR methods had better results than all the remaining methods when a small number of features was selected. However, their advantage was not so evident when the number of selected features increased. Results from the proposed mRMR methods were promising, especially in the case of the distance correlation mRMR, the kernel canonical correlation analysis mRMR and the Spearman's rank correlation mRMR. These ranking methods performed better than the original mRMR algorithm that employs mutual information internally.