Comparison of the effect of normalization techniques used inartificial neural networks on the prediction algorithm

Data set normalization process is of great importance in training complex models such as artificial neural networks obtaining effective output values. The successful results of artificial neural networks depend on the generalization ability of the network, the learning process and the effective operation of optimization algorithms and in the studies conducted, the data set normalization process has a positive effect on these situations. However, the selection of the normalization process to be applied to the data set is also a critical issue. Applying the same normalization process to each data set does not give the same performance value. There is no specific rule when choosing the appropriate normalization process for the data set. In this study, a total of 7 different normalization techniques, including minimum-maximum, d-minimum-maximum, z score, decimal, median, sigmoidal and norm normalization were applied to raw (unprocessed) data sets and a comparison was made on the effect of normalization processes on the prediction performance of artificial neural networks. The data sets used in the study were taken from the open data sets of the University of California, Irvine (UCI). As a result of the study, it was seen that normalization techniques have a positive effect on the performance of the model, but since the distributions of each data set are not the same, the normalization technique suitable for the data set also changes.