Scholarlink

<p>In industries like telecommunications and banking, scarce labeled data and high misprediction costs challenge traditional machine learning. Semi-supervised learning (SSL) and active learning (AL) offer solutions but often suffer from overconfident predictions, redundant sampling, and static pseudo-labeling thresholds. Existing methods tackle these issues separately, limiting their real-world impact. We introduce AdaptiveSSL, a scalable SSL framework that unifies Wasserstein-based uncertainty calibration, diversity-driven sampling via multi-resolution hashing, and dynamic pseudo-labeling through an adaptive, performance-driven thresholding mechanism. By iteratively refining confidence, ensuring diverse sampling, and dynamically adjusting thresholds based on a composite objective function, AdaptiveSSL achieves robust classification. On the Sparkify churn dataset, it reaches an AUC of 0.9326 with 10% labeled data, outperforming baselines by up to 10% while maintaining efficiency. On the imbalanced Home Credit dataset, it detects the minority class with an TPR score of 0.4521, surpassing model like FreeMatch. Tested on five datasets, AdaptiveSSL provides a practical tool for data-efficient, high-stakes classification.</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

AdaptiveSSL: a unified semi-supervised learning framework for robust classification via adaptive uncertainty calibration and dynamic labeling

  • 期刊论文

  • 发表日期：

    2026年1月6日

查看全文
  • Usman Gani Joy,
  • M. M. Rahman

摘要

In industries like telecommunications and banking, scarce labeled data and high misprediction costs challenge traditional machine learning. Semi-supervised learning (SSL) and active learning (AL) offer solutions but often suffer from overconfident predictions, redundant sampling, and static pseudo-labeling thresholds. Existing methods tackle these issues separately, limiting their real-world impact. We introduce AdaptiveSSL, a scalable SSL framework that unifies Wasserstein-based uncertainty calibration, diversity-driven sampling via multi-resolution hashing, and dynamic pseudo-labeling through an adaptive, performance-driven thresholding mechanism. By iteratively refining confidence, ensuring diverse sampling, and dynamically adjusting thresholds based on a composite objective function, AdaptiveSSL achieves robust classification. On the Sparkify churn dataset, it reaches an AUC of 0.9326 with 10% labeled data, outperforming baselines by up to 10% while maintaining efficiency. On the imbalanced Home Credit dataset, it detects the minority class with an TPR score of 0.4521, surpassing model like FreeMatch. Tested on five datasets, AdaptiveSSL provides a practical tool for data-efficient, high-stakes classification.