Scholarlink

<p>Currently, data stream classification has become a prominent research focus, referring to the process of labeling dynamically generated data streams. However, label scarcity and concept drift limit the effectiveness of classification models. To address these challenges, this paper proposes an instance transfer learning method leveraging clustering and local variance for drifting data streams classification, termed ITL-CLV. To reduce the distribution discrepancy between the source and target domains and enhance transferability, ITL-CLV develops a cluster-based source instance selection strategy. This strategy chooses source instances that are most similar to the target-domain distribution by calculating similarity weights and silhouette coefficients. To improve the model’s adaptability to concept drift and enhance transfer efficiency, ITL-CLV designs a novel source domain update strategy. It employs maximum posterior probability and minimum local variance methods to select high-confidence pseudo-labels for target instances, which are then used to update the model. Additionally, the minimum cross-entropy loss function is utilized to evaluate the model’s convergence. Comparative experiments with other algorithms across various data streams demonstrate that ITL-CLV achieves superior effectiveness and stability in most cases, exhibiting competitive advantages in classification accuracy and other performance metrics.</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

A novel transfer learning method leveraging clustering and local variance for drifting data streams classification

  • 期刊论文

  • 发表日期：

    2025年12月9日

查看全文
  • Jian Zhu,
  • Sanmin Liu,
  • Subin Huang,
  • Ping Zhang,
  • Wentao Yu,
  • Liwei Ye,
  • Huixian Li

摘要

Currently, data stream classification has become a prominent research focus, referring to the process of labeling dynamically generated data streams. However, label scarcity and concept drift limit the effectiveness of classification models. To address these challenges, this paper proposes an instance transfer learning method leveraging clustering and local variance for drifting data streams classification, termed ITL-CLV. To reduce the distribution discrepancy between the source and target domains and enhance transferability, ITL-CLV develops a cluster-based source instance selection strategy. This strategy chooses source instances that are most similar to the target-domain distribution by calculating similarity weights and silhouette coefficients. To improve the model’s adaptability to concept drift and enhance transfer efficiency, ITL-CLV designs a novel source domain update strategy. It employs maximum posterior probability and minimum local variance methods to select high-confidence pseudo-labels for target instances, which are then used to update the model. Additionally, the minimum cross-entropy loss function is utilized to evaluate the model’s convergence. Comparative experiments with other algorithms across various data streams demonstrate that ITL-CLV achieves superior effectiveness and stability in most cases, exhibiting competitive advantages in classification accuracy and other performance metrics.