Efficient Rule Generation for Dominant Class Problems on LARM

Lazy Associative Rule Mining (LARM) integrates lazy learning and Associative Rule Mining (ARM) to tailor label prediction results by generating related class associative rules (CARs) only when an unlabeled document comes. However, two main problems should be carefully concerned in LARM classification: (1) computing efficiency and (2) dominant class bias prediction. The main idea of the proposed method, LARM-DC, is to skip rule-inducing process in LARM so that the execution time could be greatly saved. Additionally, the confidences of LCARs are weighted to enhance rule importance on label prediction in order to correct the bias results. With regard to prediction accuracy, our experiments show that LARM-DC performs as well as LARM on balanced datasets, and gains significant improvement on imbalanced datasets. Moreover, classification efficiency is also greatly improved comparing with LARM that generally requires lots of CARs to be induced.