Is there any way to make the score.sas program produced by Enterprise Miner smaller?

[stephensloan]

The score node coming out of my SAS Enterprise Model is very large and that causes the analysis to run very slowly.  What can I do to get a smaller score.sas program output from my model?

[sbxkoenk]

Hello,

The length of the score code ( score.sas ) is only slightly related to the duration in which it can be executed.

It also depends on what exactly is done in all these statements of course.

What is your SAS Enterprise Miner model? (Gradient Boosting, SVM, regression, ...)?

The simpler the model, the faster it can be deployed and the faster it will score new data of course.

You can work with MP Connect for scoring.

Every record is scored independently from other records (which is typical for cross-sectional data).
So, you can split your data in 10 equal parts and score simultaneously in 10 remote sessions (that all run in parallel).

Here are some good introductions to MP Connect (especially the 1^st one is very clear) :

Running SAS programs in parallel using SAS/CONNECT®

By Leonid Batkhan on SAS Users January 13, 2021

https://blogs.sas.com/content/sgf/2021/01/13/running-sas-programs-in-parallel-using-sas-connect/

Base SAS + SAS/CONNECT - A simple method to generate load on any number of licensed cores

Posted 04-08-2021 05:46 AM | by SimonWilliams

https://communities.sas.com/t5/SAS-Communities-Library/Base-SAS-SAS-CONNECT-A-simple-method-to-gener...

Koen

[stephensloan]

Thank you for your reply.

I'm running an ensemble with Decision Tree and neural net.

Running in parallel is a good suggestion, but I'd still like to reduce the size of the score.sas program if possible. Are there any papers on how to do that?

[sbxkoenk]

Hello,

Allow me to say that this is a strange question.

Reducing the size of the scoring code (score.sas)??
You cannot intervene on the code itself !! That much is certain.

As far as I know, the score code purifies itself.
Meaning this --> All steps that you have done in your analysis flow that do not make it into the final model are purged automatically.

To be sure, you could make a new flow.
* Take out all variables not used by the final (ensemble) model.
* Only put nodes in your flow that take actions on the data that you really need.
With this clean process flow (which has fewer variables and fewer nodes), your score code might become simpler.

Thanks,

Koen

[Kurt_Bremser]

Reducing and optimizing code is manual work, to be done by a proficient programmer. After every step, run a regression test to verify that the code still delivers the same result.