Classification with Rejection: Concepts and Evaluations
Standard classification process allocates all processed elements to given classes. Such type of classification assumes that there are only native and no foreign elements, i.e., all processed elements are included in given classes. The quality of standard classification can be measured by two factors: numbers of correctly and incorrectly classified elements, called True Positives and False Positives. Admitting foreign elements in standard classification process increases False Positives and, in this way, deteriorates quality of classification. In this context, it is desired to reject foreign elements, i.e., not to assign them to any of given classes. Rejecting foreign elements will reduce the number of false positives, but can also reject native elements reducing True Positives as side effect. Therefore, it is important to build well-designed rejection, which will reject significant part of foreigners and only few natives. In this paper, evaluations of classification with rejection concepts are presented. Three main models: a classification without rejection, a classification with rejection, and a classification with reclassification are presented. The concepts are illustrated by flexible ensembles of binary classifiers with evaluations of each model. The proposed models can be used, in particular, as classifiers working with noised data, where recognized input is not limited to elements of known classes.
KeywordsRejection rule Binary classifiers ensemble Reclassification
The research is supported by the National Science Center, grant No 2012/07/B/ST6/01501, decision no UMO-2012/07/B/ST6/01501.
- 4.Graves, D., Pedrycz, W.: Kernel-based fuzzy clustering and fuzzy clustering: a comparative experimental study (2010)Google Scholar
- 6.Ishibuchi, H., Nakashima, T., Morisawa, T.: Voting in fuzzy rule-based systems for pattern classification problems. Fuzzy Sets Syst. 103, 223–238 (1999)Google Scholar
- 9.Luckner, M.: Comparison of hierarchical svm structures in letters recognition task. In: Rutkowski, L., Tadeusiewicz, R., Zadeh, L.A., Zurada, J. (eds.) Computational Intelligence: Methods and Applications, pp. 291–302. Challenging Problems of Science, Academic Publishing House EXIT, Warsaw (2008)Google Scholar
- 12.Pillai, I., Fumera, G., Roli, F.: A classification approach with a reject option for multi-label problems. In: Maino, G., Foresti, G.L. (eds.) ICIAP (1). Lecture Notes in Computer Science, vol. 6978, pp. 98–107. Springer (2011)Google Scholar
- 13.Schölkopf, B., et al.: Estimating the support of a high-dimensional distribution. Neural Comput. 13(7), 1443–1471 (2001)Google Scholar
- 14.Simeone, P., Marrocco, C., Tortorella, F.: Design of reject rules for ECOC classification systems. Pattern Recogn. 45(2), 863–875 (2012)Google Scholar
- 16.Xie, J., Qiu, Z., Wu, J.: Bootstrap methods for reject rules of fisher LDA. Pattern Recogn. 425–428 (2006)Google Scholar