Unexpected Short Training Time When Removing Feature Dimensions

When removing (unimportant) features to certain level, I got a suprisingly short training time, which worried us if the implementation is faulty. Could you please comment how to debug or if that is expected behavior?

The plot below shows the sudden change in training time when just removing some features. It seems the pattern is stable.


Training Parameters

BOW_XGB_init = xgb.XGBClassifier(n_estimators=100, max_depth=1, learning_rate=0.1,silent=False, objective='binary:logistic', \
                  booster='gbtree', n_jobs=32, nthread=None, gamma=0, min_child_weight=1, max_delta_step=0, \
                  subsample=1, colsample_bytree=1, colsample_bylevel=1, reg_alpha=0, reg_lambda=1)

128 physical cores; 256 GB memory; Ubuntu 22.04.2; python3.10.12


Have the model size change considerably? Try counting the average number of nodes and see if that number is different.

Thank you for the input. Each of my models is 470KB with 1000 trees and depth 1.

I also wonder why my model got 1000 trees given I set n_estimators=100. Is it because I have 10 labels for the model, so it try to have 100 trees for each label?


You can change this behavior by setting multi_strategy="multi_output_tree". See https://xgboost.readthedocs.io/en/latest/tutorials/multioutput.html for more details.