Scholarlink

<p>Data streams hide a large amount of valuable information, and it is highly important to extract this information to assist in decision-making. However, there are still many challenges and urgent problems that need to be solved, such as the scarcity of labels and concept drift. In this paper, a novel dynamic transfer learning method for drifting data streams based on cluster analysis (DCTL) is proposed to address the issues of sample labeling and concept drift in the process of data stream classification. The DCTL method consists of two important components: an instance transfer policy based on cluster analysis and a source domain updating strategy. The instance transfer policy clusters the source and target instances into the same feature space to analyze their similarity. First, to comprehensively measure the similarity weights of source instances, the number of target instances in each cluster and the relative distance from the source instances to the center of the cluster are calculated. Then, the source instances with significant similarity weights are selected to update the classifier and prepare for building a new source domain. To improve the efficiency of transfer learning and avoid negative transfer, we add an instance elimination mechanism on the basis of incremental learning in the source domain update strategy. While receiving new data for learning, source instances with low similarity weights are deleted, and new data are constantly being added to ensure that the distribution of the source domain is close to the distribution of the target domain. Through the DCTL method, the most valuable information for the target domain is selected from the source domain, and new data are continuously introduced for learning, thereby enhancing the adaptability and stability of the model. The simulation experiment compared with algorithm on five different data streams shows that DCTL is efficient and stable in most cases. In addition, DCTL maintains a competitive advantage in terms of classification accuracy and other indicators.</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

Transfer learning method based on cluster for drifting data stream classification

  • 期刊论文

  • 发表日期：

    2025年6月16日

查看全文
  • Jian Zhu,
  • Sanmin Liu,
  • Subin Huang,
  • Ping Zhang,
  • Xinquan Chen,
  • Tuyi Zhang,
  • Wentao Yu,
  • Guoyi Zhang

摘要

Data streams hide a large amount of valuable information, and it is highly important to extract this information to assist in decision-making. However, there are still many challenges and urgent problems that need to be solved, such as the scarcity of labels and concept drift. In this paper, a novel dynamic transfer learning method for drifting data streams based on cluster analysis (DCTL) is proposed to address the issues of sample labeling and concept drift in the process of data stream classification. The DCTL method consists of two important components: an instance transfer policy based on cluster analysis and a source domain updating strategy. The instance transfer policy clusters the source and target instances into the same feature space to analyze their similarity. First, to comprehensively measure the similarity weights of source instances, the number of target instances in each cluster and the relative distance from the source instances to the center of the cluster are calculated. Then, the source instances with significant similarity weights are selected to update the classifier and prepare for building a new source domain. To improve the efficiency of transfer learning and avoid negative transfer, we add an instance elimination mechanism on the basis of incremental learning in the source domain update strategy. While receiving new data for learning, source instances with low similarity weights are deleted, and new data are constantly being added to ensure that the distribution of the source domain is close to the distribution of the target domain. Through the DCTL method, the most valuable information for the target domain is selected from the source domain, and new data are continuously introduced for learning, thereby enhancing the adaptability and stability of the model. The simulation experiment compared with algorithm on five different data streams shows that DCTL is efficient and stable in most cases. In addition, DCTL maintains a competitive advantage in terms of classification accuracy and other indicators.