A method to characterize dataset based on objective rule evaluation indices

Data characterizing techniques have been developed to control learning algorithm selection by using statistical measurements of a dataset. To expand the framework of meta-learning, it is important to consider results of other learning algorithms. Therefore, we consider about a method to reuse objective rule evaluation indices of classification rules. Objective rule evaluation indices such as support, precision and recall are calculated by using a rule set and a validation dataset. This data-driven approach is often used to filter out not useful rules from obtained rule set by a rule learning algorithm. At the same time, these indices can detect differences between two validation datasets by using the rule set and the indices, because the definitions of indices independent on both of a rule and a dataset. In this paper, we present a method to characterize given datasets based on objective rule evaluation indices by using differences of correlation coefficients between each index. By comparing the differences, we describe the results of similar/dissimilar groups of the datasets.