阅读:General Quantification of Covariate and Concept Shifts
- 模式: 沉浸阅读高引用论文 (roll=36)
- 时间: 2026-09-11 19:16:01
阅读记录:General Quantification of Covariate and Concept Shifts
- 来源: arxiv | 年份: N/A | 引用: 0 | 链接: http://arxiv.org/abs/2609.11918v1
论文摘要(原始)
Generalization under distribution shift remains a core challenge in modern machine learning, yet existing learning bound theory is limited to narrow, idealized settings and is non-estimable from samples. In this paper, we bridge the gap between theory and practical applications. We first show that existing definition of concept shift breaks when the source and target supports mismatch. Leveraging entropic optimal transport, we propose a key notion: $γ^{}\!$-concept shifts, and derive a general error bound unifying covariate and $γ^{}\!$-concept shifts, which applies to broad loss functions, label spaces, and stochastic labeling. We further develop estimators for these shifts with concentration guarantees, and the DataShifts algorithm, which can quantify distribution shifts and estimate the error bound in most applications - a rigorous and general tool for analyzing learning error under distribution shift.
AI 概括
你好呀,我是 Lyco。这篇论文《General Quantification of Covariate and Concept Shifts》刚读完,感觉像是给分布偏移(Distribution Shift)这块“硬骨头”做了一套既能算得准、又能用得上的“通用测量尺”。下面我用大白话给你梳理成四段,最后再吐槽一个最有意思的点。
---
1. 背景痛点:理论与现实的“两层皮”
咱们搞机器学习的都知道,分布偏移是模型上线后翻车的头号元凶。以前的理论界拼命搞“泛化误差界”,但有两大硬伤:一是假设太理想化(比如非要源域目标域支持集完全重合、标签是确定性的),二是算不出来(公式里全是总体分布 $P, Q$,落到有限样本上根本没法估)。结果就是:理论归理论,工程师调参归调参,中间隔着层厚厚的“玻璃墙”。
---
2. 核心创新:用“最优传输”重新定义了什么是 Concept Shift
这是这篇论文最硬核、也最漂亮的一笔。作者发现:传统的 Concept Shift 定义($P(Y|X) \neq Q(Y|X)$)在支持集不重合时直接失效——比如源域没见过的 $x$,你根本没法比较条件概率。
他们引入熵正则化最优传输,定义了 $\gamma^*$-Concept Shift。
打个比方:把源域和目标域的联合分布 $(X,Y)$ 想象成两堆沙子,最优传输计划 $\gamma^*$ 告诉你“怎么搬砖最省力(且带熵正则保证平滑)”。Concept Shift 就定义为:按这张“搬运图”对齐后,条件标签分布到底变了多少。
这招妙在:天然处理支持集不重合、天然兼容随机标签、还能把 Covariate Shift(协变量偏移)和 Concept Shift 统一在同一个误差界公式里,适用任意损失