知识卡片
训练服务偏差与环境漂移分开
同一特征在训练和推理中采用不同处理会造成可避免的偏差,可复用管线降低风险。外部世界分布变化则仍可能发生,需要新数据评价;统一转换和定期重训不保证消除所有漂移。
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
- · 模型自动化测试
原书出处 · 1 处
- · 查看原书依据
1. Minimize training-serving skew. Ensure the distribution shift between training data and inference data is as small as possible, by: (i) ensuring all feature engineering logic are symmetrically applied in both scenarios (as we have done in this chapter’s code example by using scikit-learn pipelines) (ii) regularly refreshing the data that the model was trained on (more on this shortly). 2. Use production-like data for
参考来源
- 位置:《Effective Machine Learning Teams (David Tan, Ada Leung, David Colls)》第06章《模型自动化测试》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Effective Machine Learning Teams (David Tan, Ada Leung, David Colls)/_epub-src/OEBPS/ch06.html)
- 结论依据:摘录中关于训练服务偏差与环境漂移分开的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:1. Minimize training-serving skew. Ensure the distribution shift between training data and inference data is as small as possible, by: (i) ensuring all feature engineering logic are symmetrically applied in both scenarios (as we have done in this chapter’s code example by using scikit-learn pipelines) (ii) regularly refreshing the data that the model was trained on (more on this shortly). 2. Use production-like data for