Rule Generation and Rule Selection Techniques for Cost-Sensitive Associative Classification

Classification aims to assign a data object to its appropriate class, what is traditionally performed through a small dataset model such as decision tree. Associative classification is a novel strategy for performing this task where the model is composed of a particular set of association rules, in which the consequent of each rule (i.e., its right-hand-side) is restricted to the classification class attribute. Rule generation and rule selection are two major issues in associative classification. Rule generation aims to find a set of association rules that better describe the entire dataset, while rule selection aims to select, for a particular case, the best rule among all rules generated. Rule generation and rule selection techniques dramatically affect the effectiveness of the classifier. In this paper we propose new techniques for rule generation and rule selection. In our proposed technique, rules are generated based on the concept of maximal frequent class itemsets (increasing the size of the rule pattern), and then selected based on their informative value and on the cost that an error imply (possibly reducing misclassifications). We validate our techniques using two important real world problems: spam detection and protein homology detection. Further, we compare our techniques against other existing ones, ranging from well known naı̈ve-Bayes to domain-specific classifiers. Experimental results show that our techniques are able to achieve a significant improvement of 30% in the effectiveness of the classification.