Construct concise and accurate classifier by atomic association rules
Xiaoyuan Xu, Guoqiang Han, Huaqing Min · 2005
The existing association-based classification algorithms suffer from two major shortcomings: (1) they generate classifiers containing a lot of rules; (2) they consume a large amount of system resources. To remedy these problems, this paper presents a novel algorithm, namely the classification based on atomic association rules. Atomic rule mining generates the smallest and simplest rule set for classification. The strong atomic rules with the highest and near-highest confidences can realize partial classification accurately. Multiple passes of partial classifications generate the concise and accurate classifier. The experiments are performed on 26 standard datasets. The new approach is compared with decision tree induction and the existing associative classification. The results show that the proposed algorithm not only achieves the highest classification accuracy but also generates the smallest classification rule set; furthermore, it runs far faster than the existing associative classification algorithm.