知识卡片

普通读书笔记卡 · 6676.d · 生成模型:GPT-6

缩放到零一不产生校准概率

理解等级:中级

MinMaxScaler把观测分数按最小和最大值线性映射到区间,0.5仅是该区间中点。它不是类别概率校准,极端样本还会改变阈值含义;预测时应复用训练拟合的缩放,不能每批重新拟合。

这张卡片看不懂?

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

      There’s just one step left in our classification. You need to transform the score into the actual class prediction. Ideally, you’d like your score to range between 0 and 1, like a probability. Once you have the scores normalized, you can deduce the classification from the score based on a cutoff—here, we went with a simple 0.5. You can use sklearnMinMaxScaler to perform the normalization.

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第04章《从词数到语义分析》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-4.html) - 结论依据:摘录中关于缩放到零一不产生校准概率的定义、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:There’s just one step left in our classification. You need to transform the score into the actual class prediction. Ideally, you’d like your score to range between 0 and 1, like a probability. Once you have the scores normalized, you can deduce the classification from the score based on a cutoff—here, we went with a simple 0.5. You can use sklearnMinMaxScaler to perform the normalization.