This paper introduces a new diagnostic tool for understanding how trimming affects conformal prediction when calibration data is contaminated. The research analyzes fixed-threshold trimming not as a purification method, but as a conditioning process that replaces the contaminated calibration law with a retained law. The findings suggest that trimming is beneficial when anomaly scores effectively separate retention probabilities without altering the clean population, and provide templates for finite-sample certificates. AI
影响 Introduces a novel diagnostic for conformal prediction, potentially improving model reliability in the presence of noisy calibration data.
排序理由 This is a research paper published on arXiv detailing a new diagnostic for conformal prediction.
AI 生成摘要 · Google Gemini · 来自 2 个来源。 我们如何撰写摘要 →
