知识卡片
决策阈值与准确率
分类标签是离散的,模型对每个标签给分数,必须靠决策规则(如阈值:acrylic 分数超过 0.41 就预测为腈纶)把分数变成预测;阈值体现错误代价权衡,漏识别停车标志远比漏识别羊毛严重。Accuracy 是预测正确的比例,但脱离语境会误导:先要问“永远预测多数类的朴素模型有多好”。99% 对“安全过马路”可能是灾难,0.1% 对预测中奖彩票却是奇迹。听到准确率,第一问永远是“各类别的基线率是多少”;100% 更该担心过拟合或标签泄漏。
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
- Evaluating Model Validity and Quality · 查看原书依据
Because models typically report a score for a given label, such as lambs_wool: 0.6, cashmere: 0.2, acrylic: 0.1, merino_wool: 0.1, we need to invoke a decision rule of some kind to decide when we are going to predict a given label. This might involve a threshold, like “predict acrylic whenever the score for acrylic is above 0.41 for a given image,” or it might ask which class label gets the highest score out of all available o
参考来源
- 位置:《Reliable Machine Learning》第05章《Evaluating Model Validity and Quality》(源文件:extracted/Reliable Machine Learning/_epub-src/OEBPS/ch05.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:Because models typically report a score for a given label, such as lambs_wool: 0.6, cashmere: 0.2, acrylic: 0.1, merino_wool: 0.1, we need to invoke a decision rule of some kind to decide when we are going to predict a given label. This might involve a threshold, like “predict acrylic whenever the score for acrylic is above 0.41 for a given image,” or it might ask which class label gets the highest score out of all available o