One-Class Genetic Programming
One-class classification naturally only provides one-class of exemplars, the target class, from which to construct the classification model. The one-class approach is constructed from artificial data combined with the known in-class exemplars. A multi-objective fitness function in combination with a local membership function is then used to encourage a co-operative coevolutionary decomposition of the original problem under a novelty detection model of classification. Learners are therefore associated with different subsets of the target class data and encouraged to tradeoff detection versus false positive performance; where this is equivalent to assessing the misclassification of artificial exemplars versus detection of subsets of the target class. Finally, the architecture makes extensive use of active learning to reinforce the scalability of the overall approach.
KeywordsOne-Class Classification Coevolution Active Learning Problem Decomposition
Unable to display preview. Download preview PDF.
- 5.Wu, S., Banzhaf, W.: Combatting financial fraud: A coevolutionary anomaly detection approach. In: Genetic and Evolutionary Computation Conference (GECCO), pp. 1673–1680 (2008)Google Scholar
- 8.Curry, R., Heywood, M.: One-class learning with multi-objective Genetic Programming. In: Proceedings of the IEEE Systems, Man and Cybernetics Conference, SMC, pp. 1938–1945 (2007)Google Scholar
- 13.Chiu, S.: 9. In: Fuzzy Information Engineering: A Guided Tour of Applications. John Wiley & Sons, Chichester (1997)Google Scholar