A Survey on Class Imbalance Learning Algorithms in Complex Scenarios

Class imbalance introduces bias into model learning and remains a persistent and fundamental challenge in machine learning. When class imbalance is coupled with complex data distribution characteristics, such as noise interference, class overlap, and small disjuncts, these interactions significantly exacerbate the degradation of classification model performance. This paper systematically reviews class imbalance learning algorithms in complex scenarios, aiming to clarify the scope of applicability and underlying mechanisms of specific techniques. Through a critical review of the existing literature, the paper not only summarizes the latest advancements in the field, but more importantly identifies two key research gaps: 1) the limitations of current methods in addressing multi-source complexity scenarios, and 2) the ambiguous delineation of the applicability range of target algorithms. To address these issues, we design a systematic experimental framework to evaluate the performance of imbalanced classification across various complex scenarios. This framework reveals the mapping between different data complexities and algorithm effectiveness, and offers recommendations for the compatibility of techniques with specific scenarios. Finally, the paper provides guidance on matching scenarios with techniques, identifies open challenges, and outlines potential directions for future research.