The Proposal of Undersampling Method for Learning from Imbalanced Datasets

Highly imbalanced data, which occurs in many real-world applications, often makes machine-based processing difficult or even impossible. The over- and under-sampling methods help to tackle this issue, however they often have serious shortcomings. In this paper different methods of class balancing, especially those obtained by undersampling, are analyzed. Besides, a new solution is presented. The method is oriented toward finding and thinning clusters of majority class examples. Removing observations from high-density areas can lead to a less loss of information than in the case of removing individual examples or these from less-density areas. Such approach makes the distribution of examples more even. The effectiveness of the method is demonstrated through extensive comparisons to other undersampling methods with the use of eighteen public datasets. The results of experiments show that in many cases the proposed solution allows to achieve better performance than other tested techniques.