I am trying to maximize precision in a binary classification problem (there is a high cost to false positives). The data set is really unbalanced as well. Would it make sense to run a DRF or XGBOOST model twice, using the weights column the second time in order to counter-act false positives?
Are there other methods within these H2O algorithms to maximize precision (rather than log-loss) besides this potential method? I am also going to use an ensemble (which does seems to increase precision). Cross-validation does not appear to help.