Mining Transformed Data Sets

This research presents a method to select an ideal feature subset of original and transformed features. The feature selection method utilizes a genetic wrapper scheme that employs classification accuracy as its fitness function. The feature subset generated by the proposed approach usually contains features produced by different transformation schemes. The selection of transformed features provides new insight on the interactions and behaviors of the features. This method is especially effective with temporal data and provides knowledge about the dynamic nature of the process. This method was successfully applied to optimize efficiency of a circulating fluidized bed boiler at a local power plant. The computational results from the power plant demonstrate an improvement in classification accuracy, reduction in the number of rules, and decrease in computational time.