I am building a model to determine upper and lower thresholds for multiple features with respect to multiple others. (Specifically, I want to see if certain pressure measurements in an HVAC system are within normal ranges with respect to ambient temperature and electric current draw.) I would ultimately like to save such a model to PMML, or some other kind of standard.
I have had success using the quantile L1 loss in sklearn's GBT regressor as well as in LightGBM's LGBM regressor to find a single percentile for a single pressure measurement, and am able to use either sklearn2pmml or Nyoka's new pmml library to save such a model.
Is there a way of making either regressor work with multiple outputs?
Is there a way of making such a regressor work with multiple percentiles?
Is there a way to save such a pipeline to pmml or some other standard?
The first question is similar in nature to the one here, but I suspect a complete solution may look a bit different because of the PMML/exporting concerns.
