Allowing arbitrary names for setinfo in R

[hetong007]

I had a discussion with @road2stat about introducing Cox regression into xgboost. The special thing about it is that in the loss function there are some additional parameters (similar to weights).

The function setinfo enables user to set certain numerical information for each data point in an xgb.DMatrix object. In its current implementation, we only have four options for the parameter name.

@tqchen May I know the reason for that restriction? If we could remove the restriction for the names, Users could attach arbitrary parameters to the data points and it will be more flexible for the loss function implementation. This helps some applications like Cox regression.

[nanxstats]

Thanks Tom. It would be great if we could somehow relax this restriction a little. I could also do it by whitelisting new names in setinfo.xgb.DMatrix etc., but not sure if it's good practice touching the internal functions.

[khotilov]

At R level, you can create a custom objective implemented as a closure that would keep censoring or any other data associated with labels (as @tqchen replied to my suggestion at #574 ).

[nanxstats]

@khotilov thanks for the hint, do you have any example code, by any chance?

I do think using closure here will help, but I'm not too sure if that's the best way to do this at R level without complicating things unnecessarily.

[pommedeterresautee]

Here is an example of custom loss (no use of closure solution, but easy to guess how to do it)