R help - Mar 2008 - re gression trees: mean square vs. absolute errors

Hi,
I am working with CART regression now to predict a probability; the response
is binary. Could anyone tell me in which cases it is better to use mean
square error for splitting nodes and when mean absolute error should be
preferred.
I am now using the default (MSE) version and I can see that the obtained
optimal tree is very different from the tree with the least mean absolute
error.

Thanks in advance,
  Luba
-- 
View this message in context:
http://www.nabble.com/regression-trees%3A-mean-square-vs.-absolute-errors-tp16274094p16274094.html
Sent from the R help mailing list archive at Nabble.com.

You need to think through the application of your model. Is it more important
to get more cases classified correctly, or to avoid "bigger" errors
versus a
probability prediction? You should optimize your choice of a loss function
so that it is appropriate to the way in which the model will be used.


lubaroz wrote:
> 
> Hi,
> I am working with CART regression now to predict a probability; the
> response is binary. Could anyone tell me in which cases it is better to
> use mean square error for splitting nodes and when mean absolute error
> should be preferred.
> I am now using the default (MSE) version and I can see that the obtained
> optimal tree is very different from the tree with the least mean absolute
> error.
> 
> Thanks in advance,
>   Luba
> 
-- 
View this message in context:
http://www.nabble.com/regression-trees%3A-mean-square-vs.-absolute-errors-tp16274094p16286639.html
Sent from the R help mailing list archive at Nabble.com.