Abstract: In real-world scenarios, learning from data streams often faces the challenge of class imbalance, where learning algorithms are unable to effectively recognize minority class samples due to the lack of training data. To introduce the current research status and challenges of ensemble classification for imbalanced data streams, recent literature in this field is reviewed. The analysis and summary are conducted from the perspectives of decision rules based on weighting, selection, and voting, as well as learning methods based on cost-sensitive learning, active learning, and incremental learning. The performance of algorithms using the same dataset is compared. To address the imbalance issues in different types of complex data streams, ensemble classification algorithms are summarized from four aspects：concept drift, multi-class, noise, and class overlap. The time complexity of classical algorithms is analyzed. Finally, the classification challenges of imbalanced issues in dynamic data streams, data streams with missing information, multi-label data streams, and uncertain data streams are proposed for future research on ensemble strategies.

Key words: imbalanced data streams, ensemble classification, decision rule, learning methods, complex data streams

摘要： 现实世界的场景中，从数据流中学习会面临着类不平衡的问题，学习算法由于缺少训练数据而无法有效识别少数类样本。为了介绍不平衡数据流集成分类的研究现状和面临的挑战，依据近年来的不平衡数据流集成分类领域文献，从基于加权、选择和投票的决策规则和基于代价敏感学习、主动学习和增量学习的学习方式的角度详细分析和总结了不平衡数据流的集成方法，并比较了使用相同数据集的算法的性能。针对处理不同类型复杂数据流中的不平问题，从概念漂移、多类、噪声和类重叠四个方面对其集成分类算法进行总结，分析了经典算法的时间复杂度。对动态数据流、缺失信息的数据流、多标签数据流和不确定数据流中不平衡问题的分类挑战提出了下一步的集成策略研究。

关键词: 不平衡数据流, 集成分类, 决策规则, 学习方式, 复杂数据流