l*********s
发帖数: 5409 | 1 I often read claims that "xxx regression estimator has desirable features
such as shrinkage " blablabla, what is so nice about shrinkage anyway? @__@ | F****r
发帖数: 151 | | l*********s
发帖数: 5409 | 3 Two questions.
1. " shrinkage improves MSE because of extra information provided as
constraints" , if so, this means in general true regression coefficients
disperse around the origin. However, true parameters are not known to us,
nor there are anything special about the origin.
2. bias-variance trade-off . The ML estimator of mean is fairly convincing,
however, I wonder whether this is inherently linked with shrinkage. I mean,
why not having estimators with reduced variance biased away from the origin
?
【在 F****r 的大作中提到】
: Maybe this ?
: http://en.wikipedia.org/wiki/Shrinkage_estimator
| D*****a
发帖数: 2847 | 4 实质是bayesian的想法
origin没啥特别的，你可以normalize，让origin代表了你的prior belief的
期望值
比如别人做的实验，说这个参数大概是1.5。
你自己也做了些实验，如果忽略别人的结果，估计的参数是5。
如果你想把别人的结果也考虑进去，就把自己的结果向1.5的方向shrink一下。
结果就在1.5和5之间
convincing,
mean,
origin
【在 l*********s 的大作中提到】
: Two questions.
: 1. " shrinkage improves MSE because of extra information provided as
: constraints" , if so, this means in general true regression coefficients
: disperse around the origin. However, true parameters are not known to us,
: nor there are anything special about the origin.
: 2. bias-variance trade-off . The ML estimator of mean is fairly convincing,
: however, I wonder whether this is inherently linked with shrinkage. I mean,
: why not having estimators with reduced variance biased away from the origin
: ?
| l*********s
发帖数: 5409 | 5 明白了.谢谢
【在 D*****a 的大作中提到】
: 实质是bayesian的想法
: origin没啥特别的，你可以normalize，让origin代表了你的prior belief的
: 期望值
: 比如别人做的实验，说这个参数大概是1.5。
: 你自己也做了些实验，如果忽略别人的结果，估计的参数是5。
: 如果你想把别人的结果也考虑进去，就把自己的结果向1.5的方向shrink一下。
: 结果就在1.5和5之间
:
: convincing,
: mean,
|
|
