Scholarlink

<p>Conventional multi-label classification methods often fail to capture the dynamic relationships and relative intensity shifts between labels, treating them as independent entities. This limitation is particularly detrimental in tasks like sentiment analysis where emotions co-occur in nuanced proportions. To address this, we introduce a novel Weighted Difference Loss (WDL) framework. WDL operates on three core principles: (1) transforming labels into a normalized distribution to model their relative proportions; (2) computing learnable, weighted differences across this distribution to explicitly capture inter-label dynamics and trends; and (3) employing a label-shuffling augmentation to ensure the model learns intrinsic, order-invariant relationships. Our framework not only achieves state-of-the-art performance on four public benchmarks, but more importantly, it substantially improves the recognition of minority classes. This demonstrates the framework’s ability to learn from sparse data by effectively leveraging the underlying label structure, offering a robust, loss-driven alternative to complex architectural modifications.</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 weighted difference loss approach for enhancing multi-label classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年7月11日

查看全文
  • Qiong Hu,
  • Masrah Azrifah Azmi Murad,
  • Azreen Bin Azman,
  • Nurul Amelina Nasharuddin

摘要

Conventional multi-label classification methods often fail to capture the dynamic relationships and relative intensity shifts between labels, treating them as independent entities. This limitation is particularly detrimental in tasks like sentiment analysis where emotions co-occur in nuanced proportions. To address this, we introduce a novel Weighted Difference Loss (WDL) framework. WDL operates on three core principles: (1) transforming labels into a normalized distribution to model their relative proportions; (2) computing learnable, weighted differences across this distribution to explicitly capture inter-label dynamics and trends; and (3) employing a label-shuffling augmentation to ensure the model learns intrinsic, order-invariant relationships. Our framework not only achieves state-of-the-art performance on four public benchmarks, but more importantly, it substantially improves the recognition of minority classes. This demonstrates the framework’s ability to learn from sparse data by effectively leveraging the underlying label structure, offering a robust, loss-driven alternative to complex architectural modifications.