知识卡片
特征存储与训练服务偏斜
存放抽取出的特征的专用系统通常叫特征存储;即便训练系统每次都从原始数据重新抽特征,它的那些特性也依然存在。起步最快的做法是把一堆文件放在目录或对象存储的桶里,实验新特征很方便,但长期有两大缺点:一是极难保证系统整体持续正确,这类无结构的特征工程环境频繁出现训练-服务特征偏斜(同一特征在训练与服务中定义不同)以及即便只在训练系统内、特征语义随时间不一致的问题;二是妨碍协作与创新,难以追溯某特征由谁、何时、为何加入,也难以从效果好的模型反推到模型定义再到所用特征。若连出处都无法追踪,还会引发合规、伦理与法律上的治理问题。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- Training Systems · 查看原书依据
Systems like this with unstructured feature-engineering environments frequently suffer from training-serving feature skew (whereby features are defined differently in the training and serving environments) as well as problems related to inconsistent feature semantics over time, even in the training system alone. The second problem is that unstructured feature-engineering environments can actually hinder collaboration and innov
参考来源
- 位置:《Reliable Machine Learning》第07章《Training Systems》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch07.xhtml)
- 结论依据:原文明确阐述该主题,并给出定义、实践或反例。
- 原始内容:Systems like this with unstructured feature-engineering environments frequently suffer from training-serving feature skew (whereby features are defined differently in the training and serving environments) as well as problems related to inconsistent feature semantics over time, even in the training system alone. The second problem is that unstructured feature-engineering environments can actually hinder collaboration and innov