c******t 发帖数: 8 | 1 You have dataset 1, with 100 explanatory variables, response and 100
observations. You want to construct a model for prediction, but 100
variables is too many. Penalized methods, such as LASSO can be used.
The question is: you have dataset 2, with same 100 explanatory variables,
but the sample size is 1000, no response, and the data come from similar
population. How to use dataset 2 to help you construct a prediction model
for dataset 1?
Thanks | B******5 发帖数: 4676 | 2 semi-supervised learning? | R*******c 发帖数: 249 | 3 他家不会还在招吧。。。
【在 c******t 的大作中提到】![](/moin_static193/solenoid/img/up.png) : You have dataset 1, with 100 explanatory variables, response and 100 : observations. You want to construct a model for prediction, but 100 : variables is too many. Penalized methods, such as LASSO can be used. : The question is: you have dataset 2, with same 100 explanatory variables, : but the sample size is 1000, no response, and the data come from similar : population. How to use dataset 2 to help you construct a prediction model : for dataset 1? : Thanks
| G*******s 发帖数: 10605 | 4 Principal Component?
【在 c******t 的大作中提到】![](/moin_static193/solenoid/img/up.png) : You have dataset 1, with 100 explanatory variables, response and 100 : observations. You want to construct a model for prediction, but 100 : variables is too many. Penalized methods, such as LASSO can be used. : The question is: you have dataset 2, with same 100 explanatory variables, : but the sample size is 1000, no response, and the data come from similar : population. How to use dataset 2 to help you construct a prediction model : for dataset 1? : Thanks
| L*****k 发帖数: 327 | 5 this is a transfer learning problem, hehe~~many ways to do it
en, the most straightforward way is, do unsupervised learning for data1+
data2(explanatory data only) together, like dimension reduction
【在 c******t 的大作中提到】![](/moin_static193/solenoid/img/up.png) : You have dataset 1, with 100 explanatory variables, response and 100 : observations. You want to construct a model for prediction, but 100 : variables is too many. Penalized methods, such as LASSO can be used. : The question is: you have dataset 2, with same 100 explanatory variables, : but the sample size is 1000, no response, and the data come from similar : population. How to use dataset 2 to help you construct a prediction model : for dataset 1? : Thanks
| l******0 发帖数: 73 | 6 我面试时也问了这个问题,请问这是哪家的经典?
【在 R*******c 的大作中提到】![](/moin_static193/solenoid/img/up.png) : 他家不会还在招吧。。。
| o****o 发帖数: 8077 | 7 does it mean that use the information in dataset2 to prevent overfitting in
dataset1?
【在 L*****k 的大作中提到】![](/moin_static193/solenoid/img/up.png) : this is a transfer learning problem, hehe~~many ways to do it : en, the most straightforward way is, do unsupervised learning for data1+ : data2(explanatory data only) together, like dimension reduction
| s******0 发帖数: 1269 | 8 could you give more details about how to answer? | D******n 发帖数: 2836 | 9 it bears so many names.
Isnt it just like the reject-inference problem in risk modeling?
【在 B******5 的大作中提到】![](/moin_static193/solenoid/img/up.png) : semi-supervised learning?
| L*****k 发帖数: 327 | 10 in the conceptual level, if you have data from dataset2, and there is a link
between dataset 1 & 2, then the knowledge you learn from 2 should be
helpful for tasks in dataset 1
in
【在 o****o 的大作中提到】![](/moin_static193/solenoid/img/up.png) : does it mean that use the information in dataset2 to prevent overfitting in : dataset1?
| L*****k 发帖数: 327 | 11 this is not a typical semi-supervised learning problem, which assumes the
explanatory data have the same distribution
【在 D******n 的大作中提到】![](/moin_static193/solenoid/img/up.png) : it bears so many names. : Isnt it just like the reject-inference problem in risk modeling?
| w********m 发帖数: 1137 | 12 Situation: n = p -> high dimension/low power
Action: PCA -> dimension reduction
Result: avoid overfitting and heteroskedasticity. | S*x 发帖数: 705 | 13 PCA
or
Proc varclus
on the 2nd dateset
Use the selected (new) variables from 2nd dataset, build model on first
dataset
【在 c******t 的大作中提到】![](/moin_static193/solenoid/img/up.png) : You have dataset 1, with 100 explanatory variables, response and 100 : observations. You want to construct a model for prediction, but 100 : variables is too many. Penalized methods, such as LASSO can be used. : The question is: you have dataset 2, with same 100 explanatory variables, : but the sample size is 1000, no response, and the data come from similar : population. How to use dataset 2 to help you construct a prediction model : for dataset 1? : Thanks
| A*******s 发帖数: 3942 | 14 need to ask more details to see what is of interest behind that question. my
speculation is that since it is a high dimensional problem, interviewers
may expect you say something like PCA. And the task is a typical semi-
supervised learning, so maybe "first clustering, then EM" methods are
somethings they are looking for. |
|