Make a custom objective function that depends on other columns of the input data in R

#1

I am having issue implementing recency-weighting for xgboost training in R (i.e. passing a weight vector to xgb.DMatrix) - although the weighting affects the learning curve readout for the training set, it does not appear to have any impact at all on the actual model produced - performance in the test set is identical. This is very weird but
I can’t seem to get to the bottom of this issue or generate a reproducible example. So instead I would like to pass the Date column of the features to a custom loss function, something like:

custom_loss <- function(preds,dat) {
  labels <- getinfo(dat,"label")
  dates <- [a vector corresponding to the dates associated with each prediction]
  grad = f(dates)*-2*(labels - preds)
  hess = f(dates)*2
  [where f is an increasing function of the value in dates, so later samples matter more when training]
  return(list(grad=grad,hess=hess))
}

So somehow I would need to pass the vector of dates associated with whatever labels are being evaluated by the custom objective function, but I can’t seem to figure out how to do this. Any suggestions?

#2

Hmm, assigning data weights would usually affect the model produced. Unfortunately, it is currently not possible to access individual columns of a data matrix inside the custom metric function.

We should find out why changing data weights seem to have no effect. Can you post an example?

#3

I have tried to create a reproducible example but bizarrely on dummy data it does seem to work. Let me have a play around and see if I can replicate it. NB I have only weighted the dtrain matrix, not the dtest matrix.