Evaluating Difficulty of Multi-class Imbalanced Data

摘要

Multi-class imbalanced classification is more difficult than its binary counterpart. Besides typical data difficulty factors, one should also consider the complexity of relations among classes. This paper introduces a new method for examining the characteristics of multi-class data. It is based on analyzing the neighbourhood of the minority class examples and on additional information about similarities between classes. The experimental study has shown that this method is able to identify the difficulty of class distribution and that the estimated minority example safe levels are related with prediction errors of standard classifiers.