CMCCR: Classification based on multiple class-correlation rules

Previous studies show that associative classification achieves higher classification accuracy than traditional classification approaches. However, associative classification suffers from two major deficiencies and thus sometimes has lower classification accuracy. First, it often generates a large number of rules when the minimum support is set to be low. It is not only time consuming to discover so many itemsets, but also very difficult to select high quality rules from a huge set of rules. Second, the association measure confidence does not reflect correlation relationships between itemsets and the class label. To deal these problems, in this paper, we propose a new classification approach called classification based on multiple class-correlation rules (CMCCR). First, we use both support and all-confidence to mine not only frequent but also mutually associated itemsets. This method not only saves the running time but as well considerably reduces the number of itemsets generated, and thus sharply decreases the number of class-correlation rules produced. Second, we use a new correlation measure correlation confidence to discover class-correlation rules. The measure correlation confidence has two bounds:- 1and 1. It is easy to control the correlation degree of the class-correlation rules found. Finally, we use multiple class-correlation rules and average correlation degree to measure the combined effect of group rules. Experimental results on the mushroom data set show that CMCCR has higher and more stable accuracy than associative classification and decision tree method.