New to XGBoost here!
When using the python package xgboost version 1.5, we played around with setting subsample=0 and colsample_bytree=0. As expected, when subsample=0 (no training data!), performance seemed to match random classification. However, when using colsample_bytree=0 (no training attributes!), performance was unexpectedly good.
Does anyone know what xgboost is doing internally under these two circumstances? In both cases, we were surprised that no warning messages were given, as no training should be possible. We suspect that subsample is being left at 0, but that colsample_bytree may be getting reset to the default value of 1.0 — can anyone confirm/deny/shed any light on this? Thanks so much!