Abstract

In Associative classification (AC), the step of rule generation is necessarily exhaustive because of the inherited search problems from the association rule. Besides which, the entire rules set must be induced prior constructing the classifier. This article proposes a new AC algorithm called Dynamic Covering Associative Classification (DCAC) that learns each rule from a training dataset, removes its classified instances, and then learns the next rule from the remaining unclassified data rather than the original training dataset. This ensures that the exhaustive steps of rule evaluation and candidate generation will no longer be needed, thereby maintaining a real time rule generation process. The proposed algorithm constantly amends the support and confidence for each rule rather restricting itself with the support and confidence computed from the original dataset. Experiments on 20 datasets from different domains showed that the proposed algorithm generates higher quality and more accurate classifiers than other AC rule induction approaches.

Recommended Citation

Almannaee, M., S., Thabtah, F., & Lu, L. (2018). An Improved Associative Classification Algorithm based on Incremental Rules. In B. Andersson, B. Johansson, S. Carlsson, C. Barry, M. Lang, H. Linger, & C. Schneider (Eds.), Designing Digitalization (ISD2018 Proceedings). Lund, Sweden: Lund University. ISBN: 978-91-7753-876-9. http://aisel.aisnet.org/isd2014/proceedings2018/ISDMethodologies/6.

Paper Type

Event

Share

COinS
 

An Improved Associative Classification Algorithm based on Incremental Rules

In Associative classification (AC), the step of rule generation is necessarily exhaustive because of the inherited search problems from the association rule. Besides which, the entire rules set must be induced prior constructing the classifier. This article proposes a new AC algorithm called Dynamic Covering Associative Classification (DCAC) that learns each rule from a training dataset, removes its classified instances, and then learns the next rule from the remaining unclassified data rather than the original training dataset. This ensures that the exhaustive steps of rule evaluation and candidate generation will no longer be needed, thereby maintaining a real time rule generation process. The proposed algorithm constantly amends the support and confidence for each rule rather restricting itself with the support and confidence computed from the original dataset. Experiments on 20 datasets from different domains showed that the proposed algorithm generates higher quality and more accurate classifiers than other AC rule induction approaches.