Extending documentation on custom objective

[KonstantinBurkin]

This PR adds documentation to clarify how to use custom objective. Current example of custom objective is implemented using xgb.DMatrix and xgb.train (low-level version). That is not a popular way to use xgboost models. I suggest adding documentation with an example where:

• objective function is defined with numpy arrays only
• objective function is used with XGBRegressor model
• It is stated that objective for XGBRegressor usually accepts vectors in such order: (true_values, predicted_values). Whereas objective for xgb.train expects vectors in such order: (predicted_values, true_values).

[trivialfis]

Hi, the dtrain has type DMatrix, it contains the labels, along with other things like weight and qid, feature matrix. It's not the "true value", as it's a container for all data XGBoost consumes. Also, there's a section for the scikit-learn interface along with a demo custom_softmax.py.

[KonstantinBurkin]

I agree with your statement concerning dtrain type. But I could not find examples with scikit-learn interface. I want to add these examples and mention that the order of arguments is different for this case (predictions, y_test), not (dtrain, predictions).

[trivialfis]

But I could not find examples with scikit-learn interface.

https://xgboost.readthedocs.io/en/stable/tutorials/custom_metric_obj.html#scikit-learn-interface Next to the section you are modifying. Not elaborate enough, but there's a section. ;-)

order of arguments is different for this case

There shouldn't be an ordering for these parameters. They are different parameters, with different types, there's no ordering for things that are different.

[KonstantinBurkin]

I found it, thanks! Then I close my PR.