Scholarlink

<p>Active learning (AL) in open-set scenarios introduces the challenge of handling unlabeled data containing significant noise from unknown classes. Traditional entropy-based AL methods fail to handle it because they cannot distinguish between informative examples and open-set noise. To address this, we propose <b>E</b>vidence <b>CO</b>nflict (ECO) sampling, a novel query strategy to balance <i>purity</i>—prioritizing known classes and <i>informativeness</i>—maximizing learning utility. ECO leverages a naïve Bayes framework to quantify class-wise evidence and selects unlabeled examples based on two criteria: (1) <b>sufficient evidence</b> for known classes, ensuring query purity, and (2) <b>conflicting evidence</b> among classes, guaranteeing informativeness. This dual evidence-driven approach systematically filters out open-set noise while enhancing the discriminative power of the labeled dataset. We further establish a theoretical foundation for ECO, proving that the posterior variance of evidence distribution increases with a larger label space, ensuring its robustness in open-set scenarios. Extensive experiments on diverse benchmarks with varying openness ratios demonstrate the superiority of ECO over state-of-the-art methods in terms of both query efficiency and robustness to open-set noise.</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

Evidence Conflict Sampling for Open-set Active Learning

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Kun-Peng Ning,
  • Hai-Jian Ke,
  • Jia-Yu Yao,
  • Yu-Yang Liu,
  • Yong-Hong Tian,
  • Li Yuan

摘要

Active learning (AL) in open-set scenarios introduces the challenge of handling unlabeled data containing significant noise from unknown classes. Traditional entropy-based AL methods fail to handle it because they cannot distinguish between informative examples and open-set noise. To address this, we propose Evidence COnflict (ECO) sampling, a novel query strategy to balance purity—prioritizing known classes and informativeness—maximizing learning utility. ECO leverages a naïve Bayes framework to quantify class-wise evidence and selects unlabeled examples based on two criteria: (1) sufficient evidence for known classes, ensuring query purity, and (2) conflicting evidence among classes, guaranteeing informativeness. This dual evidence-driven approach systematically filters out open-set noise while enhancing the discriminative power of the labeled dataset. We further establish a theoretical foundation for ECO, proving that the posterior variance of evidence distribution increases with a larger label space, ensuring its robustness in open-set scenarios. Extensive experiments on diverse benchmarks with varying openness ratios demonstrate the superiority of ECO over state-of-the-art methods in terms of both query efficiency and robustness to open-set noise.