Skip to Content

SAP PA 3.2 code generator bug in higher polynomial models.

Hi SAP PA team,

I am trying to create various classification models through SAP PA Automated modeler but noticed some bugs in the code generator module when I tried to export code for SAP HANA backend from the tool.

When working with higher polynomial models( > 1 ) it seems that SAP PA 3.2 code generator has the below behavior, which is a bug.

1. CSV export of an auto classification model with polynomial > 1 generates wrong code, stops at first level relations and not those relations / ranges which are beyond level 1.

2. No option to export SQL code( none including HANA UDF ) of an auto classification model with polynomial > 1. Only shows CSV , C , AWK , JAVA , HTML , TeradataUDF etc.

Due to this variance, we were not able to export the generated model code from PA 3.2 to HANA.

Thanks,

Hasan

Add comment
10|10000 characters needed characters exceeded

  • Get RSS Feed

1 Answer

  • Best Answer
    Sep 15, 2017 at 08:31 AM

    Hi Hasan,
    I confirm that HANA code (SQL or UDF) is not available for models with polynomial order > 1. Not all code types are provided by the code generator, it depends on the definition of the model as per the matrix documented here.
    https://uacp2.hana.ondemand.com/viewer/74ad2cf7e54f405c8c948f00325a48e6/3.2/en-US/32ab2576a35148da8933b979d563f6ea.html

    By the way this matrix says Teradata UDF is not implemented for Order 2 models which is actually incorrect ; it is provided using a C code based implementation, not SQL.

    I will add the HANAUDF support for models with order > 1 in our product backlog. If possible I'd like to know a little more about the use case and specifically the KI increase you get when moving from order 1 to order 2 or 3, so that I can better document the feature request.

    Best Regards,
    Marc

    Add comment
    10|10000 characters needed characters exceeded

    • Hey Marc,

      Many thanks to you for your prompt answer and also including our request in the backlog. So I'll further answer, the two points requested by you:

      1. The TeraUDF output was indeed available in the polynomial 3 export model, I'd also exported the same for checking the code logic. So I concur that the documentation is wrong.
      2. Our reason for going at the poly 2 or 3 model was the F1 accuracy, which jumped from 0.794( poly 1 ) to 0.992( poly 3 ) and indeed we further validated it at next level using our validation set. In validation set the max that we could pump with level 1 model was 79.8 % and with poly 3 it started as 87 % without any other hyper-tuning.

      However we encountered that we are bound by the framework and can't export the model to HANA backend because the tool doesn't support it.

      Thanks,

      Hasan Rafiq