Scholarlink

<p>Label distribution learning is a widely studied supervised learning paradigm that effectively deals with label ambiguity by using global or local label correlation in a data-driven manner. With the growth of the dimension of the dataset, the redundancy and noise characteristics greatly weaken the effect of label distribution learning. Feature selection for spatial dimensionality reduction plays an important role in label distribution learning. Due to the complexity of label fuzziness, the traditional feature selection method that only focuses on the uniform distribution of logical labels is not suitable for label distribution learning data. Fuzzy evidence theory as a crucial approach for reasoning under uncertainty has found extensive application in various fields. In this paper, a fuzzy rough set model based on statistical distribution of data is proposed to process the label distribution learning data, and the model is combined with the fuzzy evidence theory to design a feature selection algorithm. Specifically, the fuzzy similarity between two samples in the feature space is first determined using the statistical distribution of data, integrating an adjustable parameter to regulate the similarity. The utilization of different fuzzy similarity radii leads to the establishment of a fuzzy similarity relation, ultimately enhancing the data’s classification capability. Then, a distance metric is defined in the label space, an threshold to control the closeness of two samples. According to the metric, the neighborhood class for each sample is created. Next, fuzzy belief and plausibility maps are generated by applying fuzzy evidence theory. Building upon the fuzzy belief and plausibility maps, develop two new feature selection designed for label learning data. Finally, experimental results and statistical analysis confirm that the developed algorithms effectively assess the uncertainty of label distribution learning data and achieve superior classification performance compared to four cutting-edge feature selection algorithms.</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

Feature selection for label distribution learning using fuzzy evidence theory

  • 期刊论文

  • 发表日期：

    2025年8月15日

查看全文
  • Pei Wang,
  • Hongxuan He,
  • Zhaowen Li

摘要

Label distribution learning is a widely studied supervised learning paradigm that effectively deals with label ambiguity by using global or local label correlation in a data-driven manner. With the growth of the dimension of the dataset, the redundancy and noise characteristics greatly weaken the effect of label distribution learning. Feature selection for spatial dimensionality reduction plays an important role in label distribution learning. Due to the complexity of label fuzziness, the traditional feature selection method that only focuses on the uniform distribution of logical labels is not suitable for label distribution learning data. Fuzzy evidence theory as a crucial approach for reasoning under uncertainty has found extensive application in various fields. In this paper, a fuzzy rough set model based on statistical distribution of data is proposed to process the label distribution learning data, and the model is combined with the fuzzy evidence theory to design a feature selection algorithm. Specifically, the fuzzy similarity between two samples in the feature space is first determined using the statistical distribution of data, integrating an adjustable parameter to regulate the similarity. The utilization of different fuzzy similarity radii leads to the establishment of a fuzzy similarity relation, ultimately enhancing the data’s classification capability. Then, a distance metric is defined in the label space, an threshold to control the closeness of two samples. According to the metric, the neighborhood class for each sample is created. Next, fuzzy belief and plausibility maps are generated by applying fuzzy evidence theory. Building upon the fuzzy belief and plausibility maps, develop two new feature selection designed for label learning data. Finally, experimental results and statistical analysis confirm that the developed algorithms effectively assess the uncertainty of label distribution learning data and achieve superior classification performance compared to four cutting-edge feature selection algorithms.