Same python code, the computer run, a lot of poor results come out, the feature importance is not the same, is why ah [face] monkey2:019 PNG [/face] because different seed subsample difference too much and the problem about https://github.com/dmlc/xgboost/issues/391 here