Scholarlink

<p>In multi-label learning, each instance is associated with multiple class labels rather than a single category. However, in practical scenarios, obtaining complete annotations for all training samples is often costly or infeasible, leading to missing or partially observed labels—known as the multi-label learning with missing labels (MLML) problem. Existing methods struggle to capture nonlinear feature–label relationships under incomplete supervision, resulting in degraded predictive performance. To address these limitations, a kernel alignment based algorithm for MLML named KAMLML is proposed. Motivated by the idea of aligning data representations with label structures, KAMLML introduces kernel alignment into the MLML setting, thus effectively bridging the feature and label spaces while capturing high-order nonlinear dependencies between them. To handle missing annotations, KAMLML imposes dual constraints to preserve alignment consistency with observed labels. In addition, a pseudo-label allocator is employed to dynamically infer label assignments for unlabeled samples. A proximal updating step is further incorporated to ensure smooth optimization. By leveraging the representer theorem, the learning problem is reformulated as a constrained quadratic programming task, enabling efficient optimization via standard numerical solvers. Extensive experiments on nine benchmark datasets with missing-label ratios ranging from 0.1 to 0.9 demonstrate that KAMLML consistently achieves superior performance compared with five state-of-the-art algorithms, validating its effectiveness for multi-label learning under incomplete supervision.</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

Kernel alignment based multi-label learning with missing labels

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Linlin Chen,
  • Jiwei Li,
  • Qiang He,
  • Hengyou Wang

摘要

In multi-label learning, each instance is associated with multiple class labels rather than a single category. However, in practical scenarios, obtaining complete annotations for all training samples is often costly or infeasible, leading to missing or partially observed labels—known as the multi-label learning with missing labels (MLML) problem. Existing methods struggle to capture nonlinear feature–label relationships under incomplete supervision, resulting in degraded predictive performance. To address these limitations, a kernel alignment based algorithm for MLML named KAMLML is proposed. Motivated by the idea of aligning data representations with label structures, KAMLML introduces kernel alignment into the MLML setting, thus effectively bridging the feature and label spaces while capturing high-order nonlinear dependencies between them. To handle missing annotations, KAMLML imposes dual constraints to preserve alignment consistency with observed labels. In addition, a pseudo-label allocator is employed to dynamically infer label assignments for unlabeled samples. A proximal updating step is further incorporated to ensure smooth optimization. By leveraging the representer theorem, the learning problem is reformulated as a constrained quadratic programming task, enabling efficient optimization via standard numerical solvers. Extensive experiments on nine benchmark datasets with missing-label ratios ranging from 0.1 to 0.9 demonstrate that KAMLML consistently achieves superior performance compared with five state-of-the-art algorithms, validating its effectiveness for multi-label learning under incomplete supervision.