Accuracy degradation on bosch dataset

[razdoburdin]

Hi,
I have observed the accuracy degradation for bosch-production-line-performance dataset after this PR #10298.
So, for commit ab8aa6f the accuracy is 0.844 for 100 estimators, but for ab228cc it drops to 0.762.

[trivialfis]

Could you please share your code for the benchmark? Is the accuracy from validation dataset or from the training dataset?

[trivialfis]

Initial result from my test:

* master
[99]    Train-logloss:0.02704   Train-auc:0.89945       Valid-logloss:0.03350   Valid-auc:0.71049
Train acc 0.9946429979387481
Valid acc 0.9942825668976844
Duration: 38.64595603942871

* 2.1
[99]    Train-logloss:0.02409   Train-auc:0.91872       Valid-logloss:0.03095   Valid-auc:0.71511
Train acc 0.9948367330847816
Valid acc 0.9943873189226086
Duration: 40.097050189971924