Scholarlink

Transductive few-shot learning (TFSL) is a promising approach to improve model performance in data scarcity scenarios. But the labeled samples are normally supposed to be uniformly distributed across classes, which is not only non-informative but also unrealistic. We propose an Active Learning-assisted Prototype Rectification method to address the challenges of TFSL by effectively leveraging unlabeled data and enhancing model generalization. Firstly, multi-round active learning based on the Class Distribution Difference criterion is proposed to obtain the representative and boundary-aware samples, resulting in optimal prototype initialization; Secondly, robust prototype rectification via parameter-free Laplacian smoothing enforces manifold-consistent label propagation. Our experiments demonstrate the method achieves state-of-the-art performance within the active TFSL framework while maintaining runtime efficiency close to inductive methods.

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

Efficient Transductive Few-Shot Learning with Active Learning for Imbalanced Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yujun Li,
  • Lei Yu

摘要

Transductive few-shot learning (TFSL) is a promising approach to improve model performance in data scarcity scenarios. But the labeled samples are normally supposed to be uniformly distributed across classes, which is not only non-informative but also unrealistic. We propose an Active Learning-assisted Prototype Rectification method to address the challenges of TFSL by effectively leveraging unlabeled data and enhancing model generalization. Firstly, multi-round active learning based on the Class Distribution Difference criterion is proposed to obtain the representative and boundary-aware samples, resulting in optimal prototype initialization; Secondly, robust prototype rectification via parameter-free Laplacian smoothing enforces manifold-consistent label propagation. Our experiments demonstrate the method achieves state-of-the-art performance within the active TFSL framework while maintaining runtime efficiency close to inductive methods.