When you create a SAS^® Visual Data Mining and Machine Learning project where event-based sampling is enabled, the posterior probabilities are adjusted for the prior probabilities (priors) in the score code of a modeling node. However, when you replace the data source for the project, either by using the Model Studio interface or when you retrain the project by using the downloaded project API code, the score code is adjusted incorrectly. Consequently, the predicted probabilities that are generated by the model are incorrect.

There are no errors or warnings to indicate that the priors are not adjusted in the score code.

To determine whether the priors are used, examine the code in the Node Score Code of the node results. If priors are used, then the score code should adjust the posterior probabilities like in the example below. Note that 0.8005033557 represents the prior when the target level is 0, and 0.1994966443 represents the prior when the target level is 1.

However, when the project's data source is replaced, the score code incorrectly adjusts the priors, as shown below:

Note: This issue also affects the results when you retrain your project in SAS^® Model Manager.

There is no workaround for this problem. The only way to obtain correct results is to create a new project. To speed up the process (so you do not have to rebuild your pipeline from the start), follow these steps:

1. Save the current pipeline to The Exchange as a template.
2. Create a new pipeline by selecting the saved template.

Operating System and Release Information

Viya on Linux: An update for this issue is available for SAS Viya 3.5. For instructions on how to access and apply software updates, see the Updating Your SAS Viya software section in the SAS Viya 3.5 for Linux Deployment Guide at