Pre-processing and visualisation of decision support data for enhanced machine classification

The paper reviews the pattern classification problem and examined various pre-processing strategies which could be used to simplify the classification of a given data set. In particular, the use of normalisation, mutual information scaling, rotational transforms and product features have been considered. The use of the pre-processing strategies and the data analysis tools is demonstrated using two case prediction examples. The first is data concerning a financial system and the second involved the diagnosis of faults in an electronic system. Both cases show that significant improvements in class separability are possible using quite simple and explicit pre-processing strategies. These experiments demonstrate that very significant improvements in classification accuracy possible if the attribute values in the input data applied to the perceptron are subject to mutual information scaling