Scholarlink

<p>Dynamic environments pose challenges for machine learning models due to concept drift, where data distributions change over time and may affect specific classes locally or globally. We propose One-vs-Rest Drift-Aware (OvR-DA), a framework that integrates the classical one-vs-rest decomposition with a class-informed drift detection method to enable targeted adaptation. OvR-DA monitors class-specific distributions and, upon detecting drift, selectively replaces only the affected sub-classifiers rather than retraining the entire model. This approach ensures rapid recovery and efficient handling of multi-class data streams under diverse drift scenarios while maintaining overall predictive performance. Through comprehensive experimentation and comparisons with six state-of-the-art classifiers and various One-Vs-Rest strategy variations, the proposed OvR-DA classifier consistently outperformed all non-ensemble methods and achieved competitive results with ensemble models. These findings confirm the effectiveness of OvR-DA in handling multi-class scenarios involving both local and global concept drifts. All experiments, code, and datasets are publicly accessible at <a href="https://github.com/gabrieljaguiar/multi-class-stream">https://github.com/gabrieljaguiar/multi-class-stream</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

Enhancing classification on multi-class drifting data streams with one-vs-rest strategies

  • 期刊论文

  • 发表日期：

    2026年5月7日

查看全文
  • Gabriel Jonas Aguiar,
  • Alberto Cano,
  • Juan Valentín Guerrero Cano

摘要

Dynamic environments pose challenges for machine learning models due to concept drift, where data distributions change over time and may affect specific classes locally or globally. We propose One-vs-Rest Drift-Aware (OvR-DA), a framework that integrates the classical one-vs-rest decomposition with a class-informed drift detection method to enable targeted adaptation. OvR-DA monitors class-specific distributions and, upon detecting drift, selectively replaces only the affected sub-classifiers rather than retraining the entire model. This approach ensures rapid recovery and efficient handling of multi-class data streams under diverse drift scenarios while maintaining overall predictive performance. Through comprehensive experimentation and comparisons with six state-of-the-art classifiers and various One-Vs-Rest strategy variations, the proposed OvR-DA classifier consistently outperformed all non-ensemble methods and achieved competitive results with ensemble models. These findings confirm the effectiveness of OvR-DA in handling multi-class scenarios involving both local and global concept drifts. All experiments, code, and datasets are publicly accessible at https://github.com/gabrieljaguiar/multi-class-stream.