Abstract:

In view of the problem that most of the concept drift detection algorithms have high latency and are too sensitive to noise, a concept drift detection method based on the quartile interval overlapping sliding window is proposed, which uses the samples in the quartile window and the improved Hoeffding inequality to detect the concept drift. In order to avoid the influence of noise on the classifier performance, a dynamic coefficient based on the current sample classification accuracy is introduced into the Hoeffding inequality. Experimental results show that the improved method can effectively improve concept drift detection accuracy and reduce drift detection delay.

Key words: interquartile range, Hoeffding inequality, data stream classification, concept drift

摘要：

针对大多数概念漂移检测算法都存在高延迟和对噪声过于敏感的问题，提出了一种基于四分位距交叠滑动窗口的概念漂移检测方法，该方法使用四分位距窗口中的样本和改进的Hoeffding不等式进行概念漂移检测。为更好地避免噪声对分类器性能的影响，算法在Hoeffding不等式中引入了一个基于当前样本分类正确率的动态系数。实验结果表明，改进后的方法可以有效提高概念漂移检测的准确率，减少漂移检测延迟。

关键词: 四分位距, Hoeffding不等式, 数据流分类, 概念漂移