A Novel Feature Selection Method for Regression and its Application for Prediction of Forest Fires

The current article presents a novel filter-based feature selection algorithm for regression tasks. It uses normalized mutual information to identify features relevant to a given regression task and to eliminate redundant features. The proposed method has been applied to the Forest Fires dataset. The proposed method is used to identify a set of features which is then used to fit the Support Vector Machine and perform Regression analysis. The performance of the proposed method has been compared with Sequential Feature Selection and UMIDI. The Mean Absolute Error obtained on regression analysis with the feature set identified by the proposed method is 1.106 while it is 1.11 for U-MIDI and 1.132 for Sequential Feature Selection. Thus the proposed method achieves the best performance among the three methods. It also has a smaller execution time than these methods.