知识卡片
转换特征:把计算移入特征存储以消除代码漂移
转换特征像关系数据库的存储过程:以代码写成、存放在存储系统内、对数据施加固定变换。可简单到把年龄分桶,也可组合多列或跨行计算,核心是给出一致、确定的结果。历史上这类变换常在训练阶段实现,而 serving 又要查同一特征,于是训练和 serving 各写一份,代码必然发生漂移。把它移进特征存储就消除了整类失败。若变换昂贵且列被频繁读取,存储还可物化该列,但有两个风险:定义一变就要全量重算,以及处理新数据的缺陷会让物化列与源数据不同步。
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
- Feature and Training Data · 查看原书依据
Historically, transforming features were often implemented in other locations in an ML training pipeline, commonly during the training phase. But then recall that features also need to be looked up during serving in most ML applications. So transforming features were implemented in training and in serving. And, as inevitably happens, code drift occurs. Some transforming features might be implemented slightly differently in som
参考来源
- 位置:《Reliable Machine Learning》第04章《Feature and Training Data》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch04.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:Historically, transforming features were often implemented in other locations in an ML training pipeline, commonly during the training phase. But then recall that features also need to be looked up during serving in most ML applications. So transforming features were implemented in training and in serving. And, as inevitably happens, code drift occurs. Some transforming features might be implemented slightly differently in som