Combined Cleaning and Resampling Algorithm for Multi-Class Imbalanced Data with Label Noise

The imbalanced data classification is one of the most crucial tasks facing\nmodern data analysis. Especially when combined with other difficulty factors,\nsuch as the presence of noise, overlapping class distributions, and small\ndisjuncts, data imbalance can significantly impact the classification\nperformance. Furthermore, some of the data difficulty factors are known to\naffect the performance of the existing oversampling strategies, in particular\nSMOTE and its derivatives. This effect is especially pronounced in the\nmulti-class setting, in which the mutual imbalance relationships between the\nclasses complicate even further. Despite that, most of the contemporary\nresearch in the area of data imbalance focuses on the binary classification\nproblems, while their more difficult multi-class counterparts are relatively\nunexplored. In this paper, we propose a novel oversampling technique, a\nMulti-Class Combined Cleaning and Resampling (MC-CCR) algorithm. The proposed\nmethod utilizes an energy-based approach to modeling the regions suitable for\noversampling, less affected by small disjuncts and outliers than SMOTE. It\ncombines it with a simultaneous cleaning operation, the aim of which is to\nreduce the effect of overlapping class distributions on the performance of the\nlearning algorithms. Finally, by incorporating a dedicated strategy of handling\nthe multi-class problems, MC-CCR is less affected by the loss of information\nabout the inter-class relationships than the traditional multi-class\ndecomposition strategies. Based on the results of experimental research carried\nout for many multi-class imbalanced benchmark datasets, the high robust of the\nproposed approach to noise was shown, as well as its high quality compared to\nthe state-of-art methods.\n