模型预测结果校准——Platt Scaling
方法简介:
Platt scaling is a method for transforming SVM outputs from to posterior probabilities (Platt, 1999). [1]
Platt (1999) proposed transforming SVM predictions to posterior probabilities by passing them through a sigmoid.
Platt scaling是一种参数化方法(The parametric approach), 使用LR模型(sigmoid函数)对模型的输出值进行拟合,将模型的原始输出值映射为概率值,区间(0,1)。假设f(x)为模型的输出值,
上式中的参数A,B通过在训练集(fi,yi)上进行最大似然估计获取。
适用情况:
Platt Scaling is most effective when the distortion in the predicted probabilities is sigmoid-shaped.
Platt scaling适用于模型输出的概率是S型的,this method may lead to poor calibration when the outputs do not fit the sigmoid function;
适用于样本量少的情形;
适用于SVM,boosting等算法的结果的校正。
参考文献:
[1] Alexandru Niculescu-Mizil, et al. Predicting Good Probabilities With Supervised Learning. ICML2005.
[2]vividfree的博客<使用 Isotonic Regression 校准分类器>
http://vividfree.github.io/机器学习/2015/12/21/classifier-calibration-with-isotonic-regression