Robustness of Selective Desensitization Perceptron Against Irrelevant and Partially Relevant Features in Pattern Classification
Recent practical studies have shown that a selective desensitization neural network (SDNN) is a high-performance function approximator that is robust against redundant input dimensions. This paper examined the classification performance of a single-output-SDNN, which we refer to as a selective desensitization perceptron (SDP), through a numerical experiment on binary classification problems that include some irrelevant features and partially relevant features and compared these results with multilayer perceptron (MLP) and support vector machine (SVM) classification methods. The results show that SDP was highly effective not only in dealing with irrelevant features but also in a dataset including a partially relevant feature, which is irrelevant in most of the domain but affects the output in a specific domain. These results indicate that the previously observed SDNN’s high-performance in the practical problems might be originated from the fact that SDP does not require a precise feature selection with taking account of the various degrees of feature relevance.
KeywordsBinary classification Selective desensitization perceptron Irrelevant feature Partially relevant feature
- 2.Weston, J., Mukherjee, S., Chapelle, O., Pontil, M., Poggio, T., Vapnik, V.: Feature selection for SVMs. In: Leen, T.K., Dietterich, T.G., Tresp, V. (eds.) NIPS 2000. Advances in Neural Information Processing Systems, vol. 13, pp. 647–653. MIT Press, Cambridge (2001)Google Scholar
- 6.Nonaka, K., Tanaka, F., Morita, M.: Empirical comparison of feedforward neural network on two-variable function approximation. IEICE Trans. Inf. Syst. J. 94(12), 2114–2125 (2011). (in Japanese)Google Scholar
- 7.Kawata, H., Tanaka, F., Suemitsu, A., Morita, M.: Practical surface EMG pattern classification by using a selective desensitization neural network. In: Wong, K.W., Mendis, B.S.U., Bouzerdoum, A. (eds.) ICONIP 2010. LNCS, vol. 6444, pp. 42–49. Springer, Heidelberg (2010). doi: 10.1007/978-3-642-17534-3_6 CrossRefGoogle Scholar