Preliminary Evaluation of Classification Complexity Measures on Imbalanced Data

摘要

Classification complexity measures play an important role in classifier selection and are primarily designed for balanced data. Focusing on binary classification, this paper proposes a novel methodology to evaluate their validity on imbalanced data. The twelve complexity measures composed by Ho are evaluated on synthetic imbalanced data sets with various probability distributions, various boundary shapes and various data skewness. The experimental results demonstrate that most of the complexity measures are statistically changeable as data skewness varies. They need to be revised and improved for imbalanced data.