Monotonic Constraints and Deterministic Results on PROC GRADBOOST SAS

[claudiodejesusa]

Hi everyone,

My team have some problem with PROC GRADBOOST of VDMML package on SAS Viya 3.5. I think, it is better to separate the topics.

Monotonic Constraints - We read a blog about this propertie of the proc gradboost, but we are confused about this part

" Of note for this particular data set is that the target variable "BAD" is coded as 0 and 1 with our desired predicted event level being 1, representing a bad loan. Since SAS gradient boosting defaults its event level as the first value for nominal targets (0 in our case), we will need to reverse the relationships to match our desired state."  }

Complete blog: https://blogs.sas.com/content/subconsciousmusings/2020/06/03/monotonic-constraints-with-sas/

It means that the positive event is 0 by default? It is possible to change it to 1 category? I think that's the reason, we have some non intuitive results. I would apprecciate the help of technical support of SAS.

Deterministic Results Models - We have a CAS with 1 Controller and 3 Workers in a MPP Arquitecture. We find that training models with the same person, the same data (prometed in global), and also same hyperparams, we got differents results when training and then scoring model and use some performance metric like AUC (Run 1, we got AUC 58.8 - Run 2, we got AUC 59,2, ...) . I have a hypotesis about the Arquitecture that make the result about models, but i would apprecciate that a technical of SAS or a Expert could explain that part. ¿How could it be resolved to get deterministic models?

Thanks,

Claudio

[PaigeMiller]

I think the answer is included in the part of the blog you quoted.

"we will need to reverse the relationships to match our desired state"

So, in a SAS DATA step before you run PROC GRADBOOST, in your target variable, change 0 to 1, and 1 to 0.