Session:

Neural Networks and Fuzzy Systems
Tuesday, March 02, 2004, 12.10 – 12.30

Session Chair:

Peter Anderson

Paper Title:

Perceptron Learning versus Support Vector Machines

Author(s):

Dr B. J. Falkowski, University of Applied Sciences Stralsund, Germany
S. Clausen, University of Applied Sciences Stralsund, Germany

Abstract:

In this paper the performance of perceptron learning is compared with that of support vector machines if the pertinent Vapnik-Chervonenkis (VC-) Bound is small, thus complementing some known results. To this end the creditworthiness of bank customers is evaluated using real-life (anonymous) data and a so-called scoring system. Since in practical situations there is usually a shortage of available learning data as well as a lack of sophisticated software a comparatively small number of data is used for fault tolerant perceptron learning implemented in Excel using VBA. In order to justify the tests it is shown that the VC-Bound of the raw data is indeed small. The significance of experimental results with respect to applying perceptron learning as opposed to SVMs is discussed since it allows the use of cost functions. This is of particular relevance within the banking context.

Key words: support vector machines, perceptron learning, scoring systems