Selective Pre-processing of Imbalanced Data for Improving Classification Performance

The paper discusses problems of constructing classifiers from imbalanced data. Re-sampling approaches that change the original class distribution are often used to improve performance of classifiers for the minority class. We describe a new approach to selective pre-processing of imbalanced data which combines local over-sampling of the minority class with filtering difficult examples from majority classes. In experiments focused on rule-based and tree-based classifiers we compare our approach with two other well known preprocessing methods – NCR and SMOTE. The results show that NCR is too strongly biased towards the minority class and leads to deteriorated specificity and overall accuracy, while SMOTE and our approach do not demonstrate such behavior. Analysis of a degree to which the original class distribution has been modified also reveals that our approach does not introduce so extensive changes