一种面向不平衡数据流的动态加权集成学习算法  

A dynamic weighted ensemble learning algorithm for imbalanced data stream

作　　者：江军 于化龙 JIANG Jun;YU Hualong(School of Computing,Jiangsu University of Science and Technology,Zhenjiang 212100,China)

机构地区：[1]江苏科技大学计算机学院,江苏镇江212100

出　　处：《电子设计工程》2025年第8期17-21,共5页Electronic Design Engineering

基　　金：国家自然科学基金项目资助(62176107)。

摘　　要：概念漂移现象的出现会极大影响在线学习模型的性能,特别是当数据的分布还不均衡时,模型的性能往往会受到更大的影响。针对上述问题,提出了一种面向不平衡数据流的动态加权集成增量学习算法(Incremental Dynamic Weighted Ensemble,IDWE),该算法可同时兼顾数据流中的分布不均衡和潜在的概念漂移问题。在每个新到达的数据块上,IDWE算法均利用多元高斯一类分类器对其中的少数类样本进行建模,算法在内存中维持固定数量的分类器,根据性能反馈实时更新分类器的决策权重,并根据权重变化对分类器进行动态更新。通过在6个合成数据流和2个真实数据流上进行实验,验证了算法的有效性和优越性,证明了该算法可以有效适应数据分布不均时存在的概念漂移现象。Concept drift is a major challenge in the field of online learning,and it can largely impact the performance of online learning models.Especially when data distributions are imbalanced,the impact can be further intensified.To address the issue above,we propose a dynamic weighted ensemble incremental learning algorithm called IDWE for classifying imbalanced data streams.The IDWE can simultaneously focus on the class imbalance issue and the potential concept drift issue existing in data streams.When a data chunk is received,the IDWE trains a new multivariate Gaussian one-class classifier on all minority instances contained in the data chunk.Specifically,the IDWE maintains a stable number of classifiers in its buffer and updates the weight of each of them according to performance feedback to make real-time decisions.Also,the IDWE algorithm continuously updates the buffer based on the weight variation of each classifier.Experimental results on 6 synthetic and 2 real-time data streams verify the effectiveness and superiority of the proposed algorithm,indicating that it can effectively adapt concept drift when data have imbalanced distribution.

关 键 词：概念漂移 不平衡数据流 在线学习 动态加权 集成学习 

分 类 号：TN102[电子电信—物理电子学]