q*******l
发帖数: 36 | 1 问题：
ridge regression 里面penalization那部分如果将L2 norm 换成L1 norm，那会倾向于
选择parameters更多还是更少的模型？
不是很明白这个问题是什么意思，大家帮忙看一下。
====
附：
电面，GETCO 伦敦quant trader
估计自己是没什么戏了，给需要的人参考一下吧。
面试过程非常简洁，1v1。对方法国EP出来的quant,他先自我介绍两句，我自我介绍两
句，然后直接开始问问题。
Q：GETCO 有很多products，怎样用这些作为predictor建一个revenue的模型。
A：Linear model, blabla...
Q: 你还能想到什么方法
A：nonlinear model, neural network..?
Q: predictor怎么多，会出现什么问题？
A：1.computational cost 2.不同的predictors correlated
Q: What's the name of it?
A: (ft..."name"这个单词以为是个没听过的专业术语呢，叫他重复了2遍...)
multicollinearity （单词太长了，结结巴巴说出来的...）
Q: 什么方法处理？
A：PCA，然后解释PCA
Q: 你还能想到什么方法
A：ridge regression （后悔说了，其实对RR记得不太清楚）
Q：你还能想到什么方法
A：其他factor analysis..
Q：你还能想到什么方法
A：...
Q: (叫我详细解释ridge regression, 然后就是本文开头的问题了)
A：嗯...嗯...嗯...
Q：再想想，把你思路告诉我
A：嗯...嗯...嗯...
Q：好的，没关系，谢谢，再见
（2秒之后）
Q：哦，忘了，你还有什么问题...
A：bla bla （觉得没什么戏了，也没咋问）
Q：byebye |
EM
发帖数: 715 | 2 标准的lasso, trade bias for sparsity |
q*******l
发帖数: 36 | 3 额，第一次听说这个词lasso，果然是学的不到家啊...
正在google学习中，多谢了！
【在 EM 的大作中提到】
: 标准的lasso, trade bias for sparsity
|
w**********y
发帖数: 1691 | 4 These interviewers are interesting. Given your background and their
questions, can't they even tell out the mis-match from the very beginning?
Ridge can't do variable selection. So p predictors in, p predictors out;
LASSO does. Different penalty scale gives different numbers of selected
predictors.
Mixed of ridge and LASSO, elastic net, also does. Others include SCAD, SIS,
and the old school but wrong method - stepwise.
Dimension reduction is a different topic with variable selection. This
direction includes PCA, ICA, factor model, manifold, etc.
【在 q*******l 的大作中提到】
: 额，第一次听说这个词lasso，果然是学的不到家啊...
: 正在google学习中，多谢了！
|
m**********4
发帖数: 774 | 5 大牛能否讲讲为啥old school的stepwise regression is wrong？我一直以为这种
greedy methods在现实中满实用的呢。
,
【在 w**********y 的大作中提到】
: These interviewers are interesting. Given your background and their
: questions, can't they even tell out the mis-match from the very beginning?
: Ridge can't do variable selection. So p predictors in, p predictors out;
: LASSO does. Different penalty scale gives different numbers of selected
: predictors.
: Mixed of ridge and LASSO, elastic net, also does. Others include SCAD, SIS,
: and the old school but wrong method - stepwise.
: Dimension reduction is a different topic with variable selection. This
: direction includes PCA, ICA, factor model, manifold, etc.
|
s*********y
发帖数: 284 | 6 面的是fixed income组么
【在 q*******l 的大作中提到】
: 问题：
: ridge regression 里面penalization那部分如果将L2 norm 换成L1 norm，那会倾向于
: 选择parameters更多还是更少的模型？
: 不是很明白这个问题是什么意思，大家帮忙看一下。
: ====
: 附：
: 电面，GETCO 伦敦quant trader
: 估计自己是没什么戏了，给需要的人参考一下吧。
: 面试过程非常简洁，1v1。对方法国EP出来的quant,他先自我介绍两句，我自我介绍两
: 句，然后直接开始问问题。
|
w**********y
发帖数: 1691 | 7 I meant it has theoretical flaws. To list top 3 in my mind: wrong F-test at
each step, wrong P-value, and wrong degree of freedom.
See here for more:
http://stats.stackexchange.com/questions/20836/algorithms-for-a
In situation with severes multicollinearity, both stepwise and lasso will
perform badly.
【在 m**********4 的大作中提到】
: 大牛能否讲讲为啥old school的stepwise regression is wrong？我一直以为这种
: greedy methods在现实中满实用的呢。
:
: ,
|
m**********4
发帖数: 774 | 8 Thanks! So what to do when severe multicollinearity exists? Does it mean
thatfeature selection with shrinkage methods won't work well in general?
Shall we use dimension reduction methods such as
PCA?
at
【在 w**********y 的大作中提到】
: I meant it has theoretical flaws. To list top 3 in my mind: wrong F-test at
: each step, wrong P-value, and wrong degree of freedom.
: See here for more:
: http://stats.stackexchange.com/questions/20836/algorithms-for-a
: In situation with severes multicollinearity, both stepwise and lasso will
: perform badly.
|
k*******d
发帖数: 1340 | 9 Do you mean, in Ridge case, all your betas (p in total) are still non-zero,
and in LASSO case, some beta_i will be forced to zero?
(Basically I do not exactly understand what it means by "variable selection")
,
【在 w**********y 的大作中提到】
: These interviewers are interesting. Given your background and their
: questions, can't they even tell out the mis-match from the very beginning?
: Ridge can't do variable selection. So p predictors in, p predictors out;
: LASSO does. Different penalty scale gives different numbers of selected
: predictors.
: Mixed of ridge and LASSO, elastic net, also does. Others include SCAD, SIS,
: and the old school but wrong method - stepwise.
: Dimension reduction is a different topic with variable selection. This
: direction includes PCA, ICA, factor model, manifold, etc.
|
p******i
发帖数: 1358 | |
|
|
J*****n
发帖数: 4859 | 11
哪这么复杂，去掉几个factor不就行了？
到底是人在用工具，还是工具在用人？
【在 m**********4 的大作中提到】
: Thanks! So what to do when severe multicollinearity exists? Does it mean
: thatfeature selection with shrinkage methods won't work well in general?
: Shall we use dimension reduction methods such as
: PCA?
:
: at
|
w**********y
发帖数: 1691 | 12 感觉几百年没见了？
有空出来吃饭喝酒是真理 model有鸟用 又不是t台上的model
【在 p******i 的大作中提到】
: 你们都好高端好牛逼。。。
|
w**********y
发帖数: 1691 | 13 很多东西是能手工做的 去掉几个是可以 但是可以很简单的改进为什么不，，
就像regime switch也能手动的去区别和分别估计 但是这样太糙
btw 您签名档？受什么刺激了，，
【在 J*****n 的大作中提到】
:
: 哪这么复杂，去掉几个factor不就行了？
: 到底是人在用工具，还是工具在用人？
|
w**********y
发帖数: 1691 | 14 基本就是你理解的意思
,
selection")
【在 k*******d 的大作中提到】
: Do you mean, in Ridge case, all your betas (p in total) are still non-zero,
: and in LASSO case, some beta_i will be forced to zero?
: (Basically I do not exactly understand what it means by "variable selection")
:
: ,
|
m****e
发帖数: 85 | 15 L2 norm is no greater than L1 norm, so parameters using L1 will be more
depressed.
【在 q*******l 的大作中提到】
: 问题：
: ridge regression 里面penalization那部分如果将L2 norm 换成L1 norm，那会倾向于
: 选择parameters更多还是更少的模型？
: 不是很明白这个问题是什么意思，大家帮忙看一下。
: ====
: 附：
: 电面，GETCO 伦敦quant trader
: 估计自己是没什么戏了，给需要的人参考一下吧。
: 面试过程非常简洁，1v1。对方法国EP出来的quant,他先自我介绍两句，我自我介绍两
: 句，然后直接开始问问题。
|
q*******l
发帖数: 36 | 16 Thanks a lot!
I think this was what the interviewer wanted to hear.
【在 m****e 的大作中提到】
: L2 norm is no greater than L1 norm, so parameters using L1 will be more
: depressed.
|
q*******l
发帖数: 36 | 17 yes, indeed.
【在 s*********y 的大作中提到】
: 面的是fixed income组么
|
s***e
发帖数: 267 | 18 The fact that L1 penalty leads to more sparse solution is mainly because the
subdifferential of absolute function is a set (thus zero is more likely to
be a solution).
"L2 norm is no greater than L1 norm" is not the reason as you can always
adjust the penalty coefficient to make one bigger than another or vice versa.
【在 m****e 的大作中提到】
: L2 norm is no greater than L1 norm, so parameters using L1 will be more
: depressed.
|
e**********m
发帖数: 1960 | 19 sparsity is just due to concaveness so any lp norm with p<=1 would work whil
e p>=1 would not
the
to
versa.
【在 s***e 的大作中提到】
: The fact that L1 penalty leads to more sparse solution is mainly because the
: subdifferential of absolute function is a set (thus zero is more likely to
: be a solution).
: "L2 norm is no greater than L1 norm" is not the reason as you can always
: adjust the penalty coefficient to make one bigger than another or vice versa.
|
d**0
发帖数: 124 | 20 这篇里面有讨论p的取值对problem的影响
http://arxiv.org/abs/math/0307152
whil
【在 e**********m 的大作中提到】
: sparsity is just due to concaveness so any lp norm with p<=1 would work whil
: e p>=1 would not
:
: the
: to
: versa.
|
r**a
发帖数: 536 | 21 你们有没有见过regularization term用Sobolev norm的？ |
x****r
发帖数: 17 | 22 基本不行， 首先是不sparse， 最重要的是， 在high dimension space， 你不知道那
些导数的表达式。 连domain的shape都不知道。
【在 r**a 的大作中提到】
: 你们有没有见过regularization term用Sobolev norm的？
|
