You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Indeed I think the easiest way to do it is calculating the histogram as now (for all the features of the current tree) but then picking random a subset of features and only check the reduction for these features.
In this way the histogram computing will remains exactly as now and only a random selection of features is needed at node level.
The text was updated successfully, but these errors were encountered:
My experience with xgboost is adding colsample_bynode to colsample_bytree improves a lot the performance with high noise / low signal datasets.
Time ago, in #98 issue @guolinke said about the inclusion of colsample_bylevel:
#98 (comment)
Indeed I think the easiest way to do it is calculating the histogram as now (for all the features of the current tree) but then picking random a subset of features and only check the reduction for these features.
In this way the histogram computing will remains exactly as now and only a random selection of features is needed at node level.
The text was updated successfully, but these errors were encountered: