Improving classification accuracy of associative classifiers by using k-conflict-rule preservation
Ming-Yen Lin, Tsung-Che Li, Sue-Chen Hsueh · 2013
Classification is one of the important issues in data mining. Past studies show that associative classification outperforms traditional classification techniques. The rule selection process, particularly the rule pruning mechanism, in an associative classifier generally plays an important role in classification accuracy. Most associative classifiers, such as CBA, CPAR and etc., keep only a rule of the highest confidence among conflict rules. In this paper, we propose a rule selection method called ICRP to further improve the classification accuracy. ICRP contains three rule pruning mechanisms, general-rule pruning, conflict rule pruning, and k-data coverage pruning, for preserving conflict but useful rules. The classifier has the best accuracy when the minimum support is 1%, minimum confidence is 40%, conflict threshold is 30%, and k is 2. Comprehensive experiments using the 30 well-known UCI datasets show that ICRP outperforms CPAR by approximately 7% accuracy, CBA by approximately 5.93% and CMAR by approximately 1.88%.