Scholarlink

<p>Label Distribution Learning (LDL) is a novel learning paradigm that addresses label ambiguity and has found extensive applications. While many LDL algorithms have been proposed, a common assumption is that label distributions can be represented by the maximum entropy model, which may fail to general forms of label distributions. This paper proposes a novel <b>LDL</b> method by <b>E</b>xploiting <b>N</b>eighborhoo<b>D</b>, called LDL-END, which adapts the <i>k</i>-nearest-neighbors (<i>k</i>NN) algorithm to deal with label distribution, without making any assumptions on the specific forms of label distributions. LDL-END introduces a new weight learning scheme that minimizes reconstruction error in both feature and label spaces, allowing the learned weights to depend on both. Additionally, it incorporates neighborhood label correlation to enhance performance. Theoretical analysis guarantees that the prediction of LDL-END approaches the ground-truth label distributions. Experimental results demonstrate the superior performance of LDL-END compared to state-of-the-art LDL methods.</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

Explainable label distribution learning by exploiting neighborhood

  • 期刊论文

  • 发表日期：

    2025年1月6日

查看全文
  • Nanqiao You,
  • Xiaoqiao Zhao,
  • Zhihai Gao,
  • Xuyan Song

摘要

Label Distribution Learning (LDL) is a novel learning paradigm that addresses label ambiguity and has found extensive applications. While many LDL algorithms have been proposed, a common assumption is that label distributions can be represented by the maximum entropy model, which may fail to general forms of label distributions. This paper proposes a novel LDL method by Exploiting NeighborhooD, called LDL-END, which adapts the k-nearest-neighbors (kNN) algorithm to deal with label distribution, without making any assumptions on the specific forms of label distributions. LDL-END introduces a new weight learning scheme that minimizes reconstruction error in both feature and label spaces, allowing the learned weights to depend on both. Additionally, it incorporates neighborhood label correlation to enhance performance. Theoretical analysis guarantees that the prediction of LDL-END approaches the ground-truth label distributions. Experimental results demonstrate the superior performance of LDL-END compared to state-of-the-art LDL methods.