Emerging Patterns (EPs) are item sets (characteristics) whose supports change significantly from one data class to another. This work proposes a novel approach to use EPs as a basic means for classification. It is called Bayesian Classification based on emerging patterns (BCEP). As a hybrid of the EP=based classifier and Na•ve Bayes (NB) classifier, it provides several advantages. First, it is based on theoretically well-founded mathematical models to predict an unseen case given a training sample. Second, it extends NB by using essential emerging patterns to relax the strong attribute independence assumption. Lastly, it is easy to interpret, as many unnecessary EPs are pruned based on data class coverage. An empirical study carried out on 21 benchmark datasets from the UCI Machine Learning Repository shows that our method is superior to other state-of-the-art classification methods such as C5.0, NB, CAEP and LB in terms of overall predictive accuracy.
|Cite as: Fan, H. and Ramamohanarao, K. (2003). A Bayesian Approach to Use Emerging Patterns for Classification. In Proc. Fourteenth Australasian Database Conference (ADC2003), Adelaide, Australia. CRPIT, 17. Schewe, K.-D. and Zhou, X., Eds. ACS. 39-48. |
(local if available)