A novel method for constrained class association rule mining

To create a classifier using an associative classification algorithm, a complete set of class association rules (CARs) is obtained from the training dataset. Most generated rules, however , are either redundant or insignificant. They not only confuse end users during decision-making but also decrease the performance of the classification process. Thus, it is necessary to eliminate redundant or unimportant rules as much as possible before they are used. A related problem is the discovery of interesting or useful rules. In existing classification systems, the set of such rules may not be discovered easily. However, in real world applications, end users often consider the rules with consequences that contain one of particular classes. For example, in cancer screening applications, researchers are very interested in rules that classify genes into the ''cancer'' class. This paper proposes a novel approach for mining relevant CARs that considers constraints on the rule consequent. A tree structure for storing frequent itemsets from the dataset is designed. Then, some theorems for pruning tree nodes that cannot generate rules satisfying the class constraints are provided and proved. Finally, an efficient algorithm for mining constrained CARs is presented. Experiments show that the proposed method is faster than existing methods. Association rule mining and classification are two important and common problems in the data mining field. Therefore, numerous approaches have been proposed for the integration of these models. Examples include classification based on association rules (CBA) [23], a classification model based on multiple association rules [21], a classification model based on predictive association rules [41], multi-class and multi-label associative classification [34], a classifier based on maximum entropy [35], the use of an equivalence class rule tree [39], a lattice-based approach for classification [29], the integration of taxonomy information into classifier construction [6], the integration of classification rules into a neural network [20], a condition-based classifier with a small number of rules [11], an efficient classification approach with a rule quality metric [10], and a combination of a Netconf measure and a rule ordering strategy based on rule size [15]. The implementation processes of these approaches are often similar. Firstly, a complete set of class association rules (CARs) is mined from the training dataset. Then, a subset of CARs is selected to form the classifier. However, the complete