A comparative study on rough set based class imbalance learning

This paper performs systematic comparative studies on rough set based class imbalance learning. We compare the strategies of weighting, re-sampling and filtering used in the rough set based methods for class imbalance learning. Weighting is better than re-sampling, and re-sampling is better than filtering. The weighted rough set based method achieves the best performance in class imbalance learning. Furthermore, we compare various configurations of the weighted rough set based method. The weighted rule extraction and weighted decision have greater influence on the performance of the weighted rough set based method than the weighted attribute reduction. The weighted attribute reduction based on the weighted degree of dependency, the rule extraction for the exhaustive set of rules and the weighted decision based on the majority voting of the factor of weighted strength are the optimal configurations for class imbalance learning. Finally, we compare the weighted rough set based method with the decision tree and SVM based methods. The experimental results show that the weighted rough set based method outperforms the decision tree and SVM based methods. It can be concluded from the comparisons that the weighted rough set based method is effective for class imbalance learning.