Exponential Cluster Smearing

P. Phillips, J. MacIntyre, and A. Moscardini (UK)


Neural Networks, Data Pre-processing, Clustering, Classification.


Many real world classification problems require data to be assigned to one of several classes. Classifiers in general and neural networks in particular have difficulty in classifying data points which fall between classes, or where the boundaries of two or more classes are in close proximity or overlapping. Such data will cause the neural network to become confused and the data to be assigned arbitrarily to a class. This paper describes an approach developed within the Centre for Adaptive Systems at the University of Sunderland. The Exponential Cluster Smearing (ECS) algorithm, attempts to smear data such that clusters within the data are pushed further apart from each other, while at the same time bringing data points within each cluster closer together. The result is a transformed (smeared) set of data which can be presented to a classifier, producing less confusion and better classification performance.

Important Links:

Go Back