Rule Extraction and Reduction for Hyper Surface Classification
Hyper Surface Classification (HSC), which is based on Jordan Curve Theorem in Topology, is one of the accurate and efficient classification algorithms. The hyper surface obtained by the training process exhibits excellent generalization performance on datasets not only of large size but also of high dimensionality. The classification knowledge hidden in the classifier, however, is hard to interpret by human. How to obtain the classification rules is an important problem. In this paper, we firstly extract rule from the sample directly. In order to avoid rule redundance, two optimal policies, selecting Minimal Consistent Subset (MCS) for the training set and merging some neighboring cubes, are exerted to reduce the rules set. Experimental results show that the two policies are able to accurately acquire the knowledge implied by the hyper surface and express the good generalization performance of HSC. Moreover, the time for classifying the unlabeled sample by the rules set can be shorten correspondingly.
KeywordsHSC MCS Rule Extraction Rule Reduction
Unable to display preview. Download preview PDF.
- 1.He, Q., Shi, Z.Z., Ren, L.A.: The Classification Method Based on Hyper Surface. In: Proc. 2002 IEEE Int. Joint Conference on Neural Networks, pp. 1499–1503 (2002)Google Scholar
- 4.Zhao, X.R., He, Q., Shi, Z.Z.: Hyper surface classifiers ensemble for high dimensional data sets. In: 3rd Int. Symp. Neural Networks, pp. 1299–1304 (2006)Google Scholar
- 7.Núñez, H., Angulo, C., Català, A.: Rule extraction from support vector machines. In: Proc. 2002 European Symposium on Artificial Neural Networks, pp. 107–112 (2002)Google Scholar
- 11.Cohen, W.W.: Fast effective rule induction. In: In Machine Learning: Proc. of the Twelfth International Conference, pp. 115–123 (1995)Google Scholar
- 12.Gao, B.J., Ester, M., Fraser, S., Schulte, O., Xiong, H.: The Minimum Consistent Subset Cover Problem and its Applications in Data Mining. In: Proc. the 13th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 310–319 (2007)Google Scholar