A Comparison of Two Approaches to Data Mining from Imbalanced Data

Our objective is a comparison of two data mining approaches to dealing with imbalanced data sets. The first approach is based on saving the original rule set, induced by the LEM2 (Learning from Example Module) algorithm, and changing the rule strength for all rules for the smaller class (concept) during classification. In the second approach, rule induction is split: the rule set for the larger class is induced by LEM2, while the rule set for the smaller class is induced by EXPLORE, another data mining algorithm. Results of our experiments show that both approaches increase the sensitivity compared to the original LEM2. However, the difference in performance of both approaches is statistically insignificant. Thus the appropriate approach for dealing with imbalanced data sets should be selected individually for a specific data set.