关键词: 概念漂移, 漂移检测, 数据流, 分类, Kolmogorov不等式

Abstract: In real-world data environments, data distributions often change over time, a phenomenon known as concept drift. Concept drift can significantly impact the performance of the original classification model. Therefore, when concept drift occurs, classification models need to make timely adjustments to adapt to changes in data distribution to ensure effective learning. This paper explores the potential of using the Kolmogorov inequality to detect concept drift. A Kolmogorov drift detection strategy based on error rate is proposed , a concept drift detection method is designed utilizing the Kolmogorov inequality, and the concept drift is detected abruptly or gradually in data streams. Additionally, this paper introduces a tail instance adjustment strategy, which partially mitigates the impact of old instances in the drift detection sample set, thereby further reducing drift detection delay. Experimental validation shows that compared to classical or state-of-the-art drift detectors, the proposed algorithms exhibit optimal performance in terms of classification accuracy. In terms of drift detection performance, proposed algorithms rank among the top in both false detection rate and detection delay, achieving a good balance. It demonstrates good performance in terms of runtime. Finally, proposed algorithms outperform others in overall comparison across these four metrics, meeting the expectations.

Key words: concept drift, drift detection, data stream, classification, Kolmogorov inequality