[R] Rpart, custom penalty for an error

[Maciej Bliziński]

On Sun, 2006-09-10 at 20:36 +0100, Prof Brian Ripley wrote: 
> > I am however interested in areas where the probability of success is 
> > noticeably higher than 5%, for example 20%. I've tried rpart and the 
> > weights option, increasing the weights of the success-observations.
> 
> You are 'misleading' rpart by using 'weights', claiming to have case
> weights for cases you do not have.  You need to use 'cost' instead.

As for the rpart() function, the `cost' parameter is for scaling the
variables, not for the cost of misclassifications. To specify it, the
parameter `parms' needs to be used, as a list with a `loss' element, in
form of a matrix. In other words, cost parm is not for cost, use loss
parm of the parms parm. Example usage:

tr <- rpart(y ~ x, data = some.data, method = 'class',
	parms = list(loss = matrix(c(0, 1, 20, 0), nrow = 2)))

> This is a standard issue, discussed in all good books on classification
> (including mine).

Yes, in MASS, section 12.2, Classification Theory, page 338 (fourth edition).
I was looking for it in section 9.2, where rpart() is discussed.

Thanks!

Regards,
Maciej

-- 
http://automatthias.wordpress.com