知识卡片
回归误差与对数损失
回归指标不看阈值,直接比较预测数值与真值(如预测价格)。为免误差抵消,可取绝对值得 MAE,或平方得 MSE,两者都以 0.0 为完美。MSE 重罚大误差,但含噪声或离群点时 MAE 更稳,最好两个都算,看结论是否定性不同。Log loss(源自 logit)适合把输出当真实概率:能区分 0.99、0.999、0.9999,越自信的错误代价越大;但输出恰好 1.0 或 0.0 且判错时,损失变无穷、以 NaN 毁掉均值。
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
MSE penalizes larger errors much more than smaller errors, which can be useful in domains where you do not want to make big mistakes. MSE can be less useful if the data contains noise or outlier examples that are better ignored, in which case MAE is likely a better metric.
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:MSE penalizes larger errors much more than smaller errors, which can be useful in domains where you do not want to make big mistakes. MSE can be less useful if the data contains noise or outlier examples that are better ignored, in which case MAE is likely a better metric.