Investigating the Impact of Different Representations of Data on Neural Network and Regression

In this research the impact of different data representation on the performance of neural network and regression was investigated on different datasets that has binary or Boolean class target. In addition, the performance of particular predictive data mining model could be affected with the change of data representation. The seven data representations that have been used in this research are As - Is, Min Max normalization, standard deviation normalization, sigmoidal normalization, thermometer representation, flag representation and simple binary representation. Moreover, all data representations have been applied on two datasets which are Wisconsin breast cancer and German credit dataset. As a result, the neural network performance is better than logistic regression on both datasets if we exclude the thermometer and flag representations. For datasets having a binary or Boolean target class, flag or thermometer binary representation is recommended to be used if logistic regression analysis is performed. Meanwhile, As-is representation, min max normalization, standard deviation normalization or sigmoidal normalization is recommended for neural network analysis on datasets having binary or Boolean target class.