Have tried h2o xgboost with the latest two versions (including 18.104.22.168), both give error in predictions “Categorical value out of bounds” (see attached figures for details) mainly because the categorical value when scoring is not seen during training. I have been using 22.214.171.124, which does not have this issue.
I saw a similar issue using AutoML in R language, from here. https://0xdata.atlassian.net/browse/PUBDEV-6266
Hope this can be fixed soon. Thanks!