Model Retraining- Is it using the same parameters from your model?

[sean100a]

Hello,

Im hoping somebody can put me right with this question.

I always though "model retraining" related to using the same model & parameters(variables) but it updates the hyper parameters & parameter estimates based on new data.

Is this correct for SAS Model Studio? or when you try to retrain, does Model Studio regenerate the model potentially with NEW parameters/variables that were not in the original model?

Thanks

S.

[prunellatorres]

Hey there, a newbie aboard!

Great question! Model retraining can mean different things depending on the context, but in general, it refers to updating an existing model using new data. So, your understanding is partially correct!

When you retrain a model, you typically use the same model architecture and parameters as before, but update the model's weights (or coefficients) based on new data. In this way, the model can improve its accuracy and generalizability as it learns from more examples.

As for SAS Model Studio specifically, I'm not 100% sure, but based on my experience with other machine learning tools, I would expect that it updates the existing model's weights rather than generating a completely new model. However, if there are new variables or features in the data that were not included in the original model, you may need to modify the model architecture to account for these changes.

Hope that helps! Let me know if you have any more questions.

[sean100a]

Thank you Prunella! and its certainly reassuring to know that I'm partially correct 🙂

Since posting I came across this video from SAS, its interesting as at time 6.34 it looks like model studio during re-training is running various modelling nodes and even an ensemble! This leads me to think that maybe Model Studio does introduce new variables or remove them but it'll be interesting to know definitively what it does....

https://video.sas.com/detail/video/6279090663001/modelops-governance-%7C-part-3:-monitoring-and-retr...

It'll be good to hear what the folks from SAS think!

[BrettWujek · Accepted solution]

You are correct. For registered models that came from a Model Studio project, "retrain" actually reruns the pipelines in the project and may return an entirely different model as the new champion for the project. In a sense, consider a Model Studio project as the machine that outputs a model - so if you give it new data it will provide a new model (not just a retrained original champion model).

That being said, I can definitely see the need here to "lock down" the model in use, at least from the standpoint of which model type and which hyperparameters are used, and just retrain that individual model. So I will create a new Requirement for us to add to our release planning at some point.

[sean100a]

Thanks so much for the detailed explanation and the quick response.

Yes i think there are definitely situations where restricting it to just retrain the current model parameters would help.