Scholarlink

<p>Concept drift in non-stationary data streams poses challenges to machine learning models by degrading predictive performance over time. Although unsupervised detectors resolve the practical limitation of label unavailability, many current approaches rely on fully adaptive frameworks where both the statistical reference and feature representation evolve continuously. This reliance on fluid baselines can lead to representational erosion, where the detector gradually loses its memory of the original concept, making it desensitized to significant, abrupt deviations. To address this deficiency, we propose Concept Anchor Drift Detection (CADD), a novel unsupervised framework. CADD establishes a stable semantic baseline by training a deep autoencoder a single time on an initial, trusted segment of the data stream. The resulting frozen encoder serves as a constant “Concept Anchor,” projecting all subsequent data into a fixed, semantically meaningful latent space. Drift is then detected by applying a dimension-wise Kolmogorov-Smirnov test to a sliding window of these anchored latent vectors. This design balances stability (via the fixed encoder) with plasticity (via the sliding window), preventing erosion while allowing adaptation to new valid concepts. A broad empirical evaluation on real-world and synthetic benchmarks demonstrates that CADD achieves highly competitive predictive performance, particularly excelling among single-model detectors in scenarios characterized by abrupt and significant distributional shifts. While acknowledging trade-offs in environments dominated by gradual or seasonal changes, CADD significantly outperforms adaptive baselines on abrupt drift benchmarks, validating the methodological value of a fixed representational anchor and introducing a robust, proactive tool for maintaining model reliability in evolving environments.</p>

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

Concept anchor drift detection: a framework for proactive detection of abrupt drifts in real-world data streams

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Meysam Roostaee,
  • Mohammad Hosein Soheilian

摘要

Concept drift in non-stationary data streams poses challenges to machine learning models by degrading predictive performance over time. Although unsupervised detectors resolve the practical limitation of label unavailability, many current approaches rely on fully adaptive frameworks where both the statistical reference and feature representation evolve continuously. This reliance on fluid baselines can lead to representational erosion, where the detector gradually loses its memory of the original concept, making it desensitized to significant, abrupt deviations. To address this deficiency, we propose Concept Anchor Drift Detection (CADD), a novel unsupervised framework. CADD establishes a stable semantic baseline by training a deep autoencoder a single time on an initial, trusted segment of the data stream. The resulting frozen encoder serves as a constant “Concept Anchor,” projecting all subsequent data into a fixed, semantically meaningful latent space. Drift is then detected by applying a dimension-wise Kolmogorov-Smirnov test to a sliding window of these anchored latent vectors. This design balances stability (via the fixed encoder) with plasticity (via the sliding window), preventing erosion while allowing adaptation to new valid concepts. A broad empirical evaluation on real-world and synthetic benchmarks demonstrates that CADD achieves highly competitive predictive performance, particularly excelling among single-model detectors in scenarios characterized by abrupt and significant distributional shifts. While acknowledging trade-offs in environments dominated by gradual or seasonal changes, CADD significantly outperforms adaptive baselines on abrupt drift benchmarks, validating the methodological value of a fixed representational anchor and introducing a robust, proactive tool for maintaining model reliability in evolving environments.