Improving the Classification Ability of Neural Networks using Exponential Smearing on Clustered Data

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


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


A common problem in using neural networks for classification purposes is the complex nature of the data, which may be of high dimensionality and have convoluted and overlapping classes which are unknown to the user. As a consequence classification performance of the neural network suffers dramatically. This paper describes the development of a technique which improves the classification ability of neural networks by pre-processing pre-clustered data using an exponential smearing algorithm. The result is smeared data which are easier for the neural network to separate into the appropriate classes based on whatever clustering technique is used. The exponential smearing algorithm is shown to work in two, three and four dimensions, and the paper presents results of work on benchmark data (eg the Iris problem).

Important Links:

Go Back