Abstract: Many studies have shown that associative classification is a promising classification method．However,most algorithms of associative classifications may not achieve high classification performance on imbalanced datasets because they generate rules based on the “support-confidence” framework．The confidence (support) tends to bias the majority class in imbalanced datasets．As a result,these instances with minority class may be misclassified．We proposed a new associative classification approach called CRAC (Correlated Rules based Associative Classification for Imbalanced Datasets)．First,we mine frequent and mutual associative itemsets for classification．Therefore,we will generate small set of high-quality rules．Second,CRAC only select the rule with largest lift as a CAR among all rules with that frequent and associative itemset as condition．As a result,the antecedent and the consequent of the rules CRAC generated are positively correlated．Finally,we rank rules according to a new metric which integrates lift,support and Complement Class Support (CCS)．So,we are likely to use rules with positively correlation to prediction the minority class．Our experiments on fifteen UCI data sets show that our approach is an effective classification technique for both balance and imbalanced datasets,and has better average classification accuracy in comparison with CBA.

Key words: Data mining,Associative classification,Imbalance datasets,Correlated rules