A Study of Predictive Accuracy for Four Associative Classifiers

Association rule discovery is one of the primary tasks in data mining that extracts patterns to describe correlations between items in a transactional database. Using association rule mining for constructing classification systems is a promising approach. There are many associative classification approaches that have been proposed recently such as CBA, CMAR and MCAR. In this research paper, four associative rule algorithms (CBA, CMAR, CPAR, MCAR) have been compared with reference to accuracy against 12 benchmark classification problems. Our goal is to determine the most accurate technique in forecasting the future classes of unseen test data objects. After experimentation with different data sets, the results revealed that none of the investigated techniques dominated the others with regards to accuracy. Moreover, MCAR produced more accurate classification systems than CBA, CMAR and CPAR, respectively. This is due to the less pruning operation employed by MCAR, which leads to generating larger classifiers. A post pruning method is recommended to reduce the number of rules generated by MCAR, where it is obvious for cases like “Cleve” and “Germany” data sets.