基于可信多数投票的快速概念漂移检测

数据流因具有数据持续到达，概念漂移产生时刻无法预测、概念的数量不确定等特征，使得滑动窗口的大小很难事先确定，滑动窗口包含概念的数量对概念漂移检测存在影响.本文提出了基于可信多数投票的快速概念漂移检测算法（CMV_SEA），该算法使用SEA算法中的基分类器淘汰方法，使用可信多数投票实现滑动窗口中基分类器的集成.仿真实验表明：相比于SEA算法，CMV_SEA算法提高了泛化能力；能在新概念产生的第一时间内检测到概念漂移；对概念漂移的检测能力和新概念的学习能力不受滑动窗口大小的影响.

Fast Detection of Concept Drifts Based on Confident Majority Voting

Data stream is characterized by the continuous arrival of data, the unpredictability of the time of concept drifts and the uncertain number of concepts, making it difficult to predict the size of the window in advance, but the number of concepts in the window affects the detection of concept drifts. This paper proposed an algorithm for the fast detection of concept drifts in data streams by taking a confident majority voting strategy (CMV_SEA). This algorithm replaced base classifiers in a window like SEA and used majority voting strategy to ensemble all base classifiers in the window. The experiment results illustrate that CMV_SEA can promote prediction accuracy and detect concept drifts as soon as a new concept comes up, and its ability to detect and learn a new concept is not influenced by the size of the window.