Potential Anchoring for imbalanced data classification

• Proposal of potential resemblance loss for measuring relative class distribution shape. unified over and undersampling framework based on resemblance. data difficulty index evaluation dataset complexity. Experimental the proposed approach. Examination factors influencing performance Data imbalance remains one negatively affecting contemporary machine learning algorithms. One most common approaches to reducing negative impact is preprocessing original with data-level strategies. In this paper we propose a imbalanced over- undersampling. The approach utilizes radial basis functions preserve shape underlying distributions during resampling process. This done by optimizing positions generated synthetic observations respect loss. final Potential Anchoring algorithm combines within framework. results experiments conducted 60 datasets show outperformance state-of-the-art algorithms, including previously methods that utilize model potential. Furthermore, analysis complexity particularly well suited handling naturally complex (i.e. not affected presence noise) datasets.