Abstract：Instances in some classes are rare in multiclass imbalanced datasets and therefore few rules for these classes are generated by support-confidence based associative classification algorithms. Consequently, instances in these minority classes are difficult to be correctly classified. Aiming at this problem, an improved associative classification algorithm for multiclass imbalanced datasets is proposed. To extract more rules for minority classes, rules are extracted according to positive correlation between itemsets and classes. Then, to improve the priority of minority classes rules, the rule strength based on itemsets class distribution is designed to rank rules. Finally, to address problems of no matched rules or matched rules in conflict, a k nearest neighbor algorithm is incorporated into the improved associative classification to classify new instances. Experimental results show that the proposed algorithm extracts more minority classes rules and promotes the priority of the minority classes rules compared with support-confidence based associative classification, and thus G-mean and F-score value for multiclass imbalance datasets are improved.
2015, Vol. 28 
