Scholarlink

<p>Dynamic environments pose significant challenges to machine learning models due to the presence of concept drift, wherein the underlying data distribution evolves over time. Moreover, concept drift can affect only a subset of classes globally or locally. To address this, we propose the Class Informed Drift Detection Method (CIDDM), a novel and classifier-agnostic approach for detecting concept drift in multi-class data streams. CIDDM monitors class-specific data distributions using multiple sliding windows and statistical descriptors—such as mean, kurtosis, and eigenvalues—to identify deviations from learned concepts. Through rigorous and extensive experimentation and comparison with 11 state-of-the-art methods, the proposed drift detector demonstrates superior performance in detecting concept drifts, particularly in scenarios with local changes within the data stream. Moreover, CIDDM’s classifier-agnostic nature ensures its adaptability across diverse learning models. Furthermore, we investigate the impact of combining drift detectors and classifiers on predictive performance. The results of this study validate CIDDM’s efficacy in detecting concept drift and therefore enhancing the resilience of machine learning systems to concept drift, thereby advancing the field of stream learning and bolstering the reliability of machine learning applications in dynamic environments. Experiments, code and data are publicly available at <a href="https://anonymous.4open.science/r/class-informed-drift-detector-B396/">https://anonymous.4open.science/r/class-informed-drift-detector-B396/</a>.</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

Class Informed Concept Drift Detection in Multi-Class Data Streams

  • 期刊论文

  • 发表日期：

    2025年12月2日

查看全文
  • Gabriel Jonas Aguiar,
  • Alberto Cano

摘要

Dynamic environments pose significant challenges to machine learning models due to the presence of concept drift, wherein the underlying data distribution evolves over time. Moreover, concept drift can affect only a subset of classes globally or locally. To address this, we propose the Class Informed Drift Detection Method (CIDDM), a novel and classifier-agnostic approach for detecting concept drift in multi-class data streams. CIDDM monitors class-specific data distributions using multiple sliding windows and statistical descriptors—such as mean, kurtosis, and eigenvalues—to identify deviations from learned concepts. Through rigorous and extensive experimentation and comparison with 11 state-of-the-art methods, the proposed drift detector demonstrates superior performance in detecting concept drifts, particularly in scenarios with local changes within the data stream. Moreover, CIDDM’s classifier-agnostic nature ensures its adaptability across diverse learning models. Furthermore, we investigate the impact of combining drift detectors and classifiers on predictive performance. The results of this study validate CIDDM’s efficacy in detecting concept drift and therefore enhancing the resilience of machine learning systems to concept drift, thereby advancing the field of stream learning and bolstering the reliability of machine learning applications in dynamic environments. Experiments, code and data are publicly available at https://anonymous.4open.science/r/class-informed-drift-detector-B396/.