Score mismatch while using JPMML to evaluate a SAS produced PMML file

[yogesh927]

I have a PMML generated from SAS Miner that I can't get properly evaluated using JPMML 1.2.6. i am getting a significant difference in scores when comparing with SAS.

Earlier,I was getting an FMTWIDTH error and resolved it using the solution mentioned in

http://stackoverflow.com/a/33157862/1808924

Any ideas why I am not getting same score using JPMML?

[JasonXin]

yogesh927, First, thanks for using SAS. My name is Jason Xin, advanced analytics solution architect working at SAS Institute. SAS Enterprise Miner 14.1 PMML is updated to full compatibility with DMG Version 4.1. Different versions of Enterprise Miner have different DMG version to match. Since you are able to use Enterprise Miner, you should have access to its in-product document. You can search for PMML there. It should tell you which DMG version THE EM version you are running is compatible with. Challenge is DMG versions are not backward compatible. Therefore a newer version of EM such as 14.1 that is compatible with DMG 4.1 may still have problems with external PMML scoring engines that are not specifically DMG 4.1 compatible. I Googled for JPMML 1.2.6. I would directionally recommend to make sure JPMML 1.2.6 is fully compatible with the DMG version your EM is comfortable with. Good luck. Jason Xin

[yogesh927]

Thanks Jason for a detailed reply.

>  I would directionally recommend to make sure JPMML 1.2.6 is fully compatible with the DMG version your EM is comfortable with.

I have verified this by generating PMML DMG Version 4.1. using R studio and getting exact results with JPMML.But i am not sure why in case of SAS generated PMLL it's not working fine .

Although, researching  a bit more on the issue i found a pattern that the scoring probabilities JPMMl is calculating is exactly double of SAS Score.please see the attached file.

[JasonXin]

Perhaps you can open a technical support ticket with SAS and have them review the details. I am not very good at XML details. Sorry. Jason Xin