Abstract
Learning systems often describe a target class as a disjunction of conjunctions of conditions. Recent work has noted that small disjuncts, i.e., those supported by few training examples, typically have poor predictive accuracy. One model of this accuracy is provided by the Bayes-Laplace formula based on the number of training examples covered by the disjunct and the number of them belonging to the target class. However, experiments show that small disjuncts associated with target classes of different relative frequencies tend to have different error rates. This note defines the context of a disjunct as the set of training examples that fail to satisfy at most one of its conditions. An empirical adaptation of the Bayes-Laplace formula is presented that also makes use of the relative frequency of the target class in this context. Trials are reported comparing the performance of the original formula and the adaptation in six learning tasks.
Article PDF
Similar content being viewed by others
References
BreimanL., FriedmanJ.H., OlshenR.A. & StoneC.J. (1984) Classification and regression trees. Belmont, CA: Wadsworth.
HolteR.C., AckerL.E., & PorterB.W. (1989). Concept learning and the accuracy of small disjuncts. Proceedings of the 11th International Joint Conference on Artificial Intelligence (pp. 813–818). Detroit, MI: Morgan Kaufmann.
KruskalW.H., & TanurJ.M. (1978). International encyclopedia of statistics. New York, NY: Free Press.
NiblettT. (1987). Constructing decision trees in noisy domains. In I.Bratko and N.Lavrač (Eds.), Progress in machine learning. Winslow, U.K.: Sigma Press.
QuinlanJ.R., ComptonP.J., HornK.A., &LazarusL. (1987). Inductive knowledge acquisition: A case study. In J.R.Quinlan (Ed.), Applications of expert systems. Wokingham, U.K.: Addison Wesley.
Author information
Authors and Affiliations
Rights and permissions
About this article
Cite this article
Quinlan, J.R. Improved estimates for the accuracy of small disjuncts. Mach Learn 6, 93–98 (1991). https://doi.org/10.1007/BF00153762
Issue Date:
DOI: https://doi.org/10.1007/BF00153762