Defining a callback to write hessians of train observations to a csv file

[jaguerrerod]

I would like to inspect the distribution of gradients and hessians during training to understand the impact of the min_child_weight parameter on the split process. Is there a way to define a callback for this using the R interface?

[trivialfis]

I think you need to define a custom objective for that.

[jaguerrerod]

That's the way when it's possible to define the gradient and hessian within a function, but my issue is with the use of rank:pairwise, where I cannot calculate them within a customizable objective function.
I have no idea about the ranges in which gradients and hessians fluctuate, and therefore, I'll have no choice but to try the regularization parameters alpha, lambda, and min_child_weight with a logarithmic scale to efficiently cover a wide range.
It would be interesting to have certain objects available to use within the callbacks, such as in this case the array of gradients and hessians.
Another idea I've been considering for some time is to use a split selection criterion based on efficiency, defined not only by gain but also by how the cover of the parent node is divided by said split.
In the case of two similar gains, I prefer to split the node into two parts of the most similar size, so the imbalance of the sizes of the child nodes can be included as a penalty to the gain.
This is interesting when the signal in the data is very small and the noise is large, so the estimates of the values of the leaves with small size are unstable. This strategy complements the use of relatively large min_child_weight.
To experiment with this, I would need information about the candidates for split points along with their respective gains and the cover of the child nodes they generate.
Anyway, for now, I'll focus on optimizing the available parameters, but enabling more information for the callback could be an interesting way of experimentation without needing to modify the code.