Classification by Association Rules: The Importance of Minimal Rule Sets

Building a classification system using datamining techniques has shown lower error rates than traditional algorithms such as decision trees. However, because the number of possible association rules in general is very large, algorithms are usually complicated and prone to overfitting. In this paper, we present a new algorithm that generates and uses a minimum set of association rules to form the classifier. Experiments show this algorithm obtains lower error rates than other two popular algorithms in 17 of 26 benchmark datasets, among which 6 to 8 have confidence levels of more than 95%.