Scholarlink

As concept drifts change the feature of data streams, detecting concept drifts is significant for analyzing data streams. ABCD (Adaptive Bernstein Change Detector) is a drift detection method based on autoencoders and is one of the most accurate drift detectors. It learns the current property of a given stream by training an autoencoder (AE). Then, it alerts a concept drift when the reconstruction errors of AE grow large for recent data. To adapt to the new concept, ABCD updates the AE whenever a concept drift happens. This paper shows that, even without updating the AE, we can create a more precise drift detector than ABCD by coupling the non-updated autoencoder-based drift detector with another simple lightweight one.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Accurate Concept Drift Detection Without Updating Autoencoders

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Taisei Takano,
  • Hisashi Koga

摘要

As concept drifts change the feature of data streams, detecting concept drifts is significant for analyzing data streams. ABCD (Adaptive Bernstein Change Detector) is a drift detection method based on autoencoders and is one of the most accurate drift detectors. It learns the current property of a given stream by training an autoencoder (AE). Then, it alerts a concept drift when the reconstruction errors of AE grow large for recent data. To adapt to the new concept, ABCD updates the AE whenever a concept drift happens. This paper shows that, even without updating the AE, we can create a more precise drift detector than ABCD by coupling the non-updated autoencoder-based drift detector with another simple lightweight one.