Better handling for categorical variables

[AbdealiLoKo]

This is a feature request.

Decision trees lend themselves naturally to categorical variables. It would be nice if xgboost could handle categorical variables inherently. It could read the feature-map or something to identify categorical variables and try to handle them in a better way.

This is important because sometimes it is difficult to encode these categorical variables into numerical values. For large data, One-Hot-Encoding gives waay too many buckets making the data blow up in size. And ordinal encoding assumes a implicit order in the data which is not obvious - for example: airport code.

Many other types of encoding do exist, but why do a round-about way if decision trees already lend themselves to categorical variables more naturally than numerical variables !?

[pommedeterresautee]

The politic of XGBoost is to not have a special support for categorical variables. It s up to you to manage them before providing the features to the algo.