Stop training branch of tree once a specific feature is used

[ghost]

I am doing some multivariate time series experiments with only partially stationary data. The idea is this:

1. First I introduce an additional feature which is monotonically increasing (but not strictly), e.g. an integer year_index assuming that I have several years of data.
2. Then I want to train a gradient boosting model. However, as soon as the feature year_index would normally be used, I want to leave the current branch of the tree unchanged.

The reasoning behind this idea is that if there is no better feature than year_index, then we have already exploited all information which is useful for generalization of future data.

Is it possible to achieve such an effect without changing the C++ code, e.g. by modifying a tree immediately after it has been created in a callback or with paramaters?

[StrikerRUS]

I suppose the described behavior can be achieved via a callback in which you dump a current model and search for a specific (allowed to be used only once) feature. Then you stop training in case of a successful search.

However, I'm going to add this issue into our feature requests hub for neater out of the box solution. Let's see how demanded this feature is.

[StrikerRUS]

Closed in favor of being in #2302. We decided to keep all feature requests in one place.

Welcome to contribute this feature! Please re-open this issue (or post a comment if you are not a topic starter) if you are actively working on implementing this feature.