Efficient and Effective Feature Selection in the Presence of Feature Interaction and Noise
This paper addresses the problem of feature subset selection for classification tasks. In particular, it focuses on the initial stages of complex realworld classification tasks when feature interaction is expected but illunderstood, and noise contaminating actual feature vectors must be expected to further complicate the classification problem. A neural-network based featureranking technique, the ‘clamping’ technique, is proposed as a robust and effective basis for feature selection that is more efficient than the established comparable techniques of sequential floating searches. The efficiency gain is that of an Order(n) algorithm over the Order(n 2) floating search techniques. These claims are supported by an empirical study of a complex classification task.
KeywordsFeature Selection Classification Accuracy Feature Subset Feature Interaction Feature Subset Selection
Unable to display preview. Download preview PDF.
- 3.Wang, W., Jones, P., and Partridge, D. A comparative study of feature salience ranking techniques, Neural Computation, 2000 (in press).Google Scholar
- 5.Jain A, Chandrasekaran B. Dimensionality and sample size considerations. In: Krishnaiah P R, Kanal L N (eds.) Pattern Recognition in Practice. North Holland, 1982, vol. 2, chap. 39, pp. 835–855Google Scholar
- 6.Theodoridis S, Koutroumbas K. Pattern Recognition, Academic Press, San Diego, 1999Google Scholar
- 8.Mao J, Mohiuddin K, Jain A K. Parsimonious network design and feature selection through node pruning, Proc. 12th ICPR, Jerusalem, pp. 622–624, 1994Google Scholar
- 9.Rumelhart D E, Hinton G E, Williams R J. Learning internal representations by error propagation. In: Rumelhart D E, McClelland J L (eds) Parallel Distributed Processing: Explorations in the Microstructure of Cognition. MIT Press, Cambridge, Mass., 1986, pp. 318–362Google Scholar