知识卡片
欠采样减少成本也会丢覆盖
减少多数类可缓解训练失衡,却可能丢掉重要变式和困难负例。数量越少不必越好,需比较保留策略与目标误差;训练比例改变后概率和决策阈值也可能需要重新校准。
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
- · 查看原书依据
Undersampling is the opposite side of the same coin. With this technique, you just drop examples from the overrepresented class. In the dog–cat example, we would randomly drop 19,800 cat images and be left with 400 examples, half dog and half cat. This, of course, causes a glaring problem of its own: we’ve thrown away the vast majority of the data and are working from
参考来源
- 位置:《Natural Language Processing in Action, Second Edition》附录D《附录D:机器学习工具与技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/appendix-d.html)
- 结论依据:摘录中关于欠采样减少成本也会丢覆盖的定义、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:Undersampling is the opposite side of the same coin. With this technique, you just drop examples from the overrepresented class. In the dog–cat example, we would randomly drop 19,800 cat images and be left with 400 examples, half dog and half cat. This, of course, causes a glaring problem of its own: we’ve thrown away the vast majority of the data and are working from