Adaptive Model learning while migrating to a different class
We are migrating our application from 7.3.1 Pega marketing to a 8.1.4 version. Due to some architectural decisions from the client side we had to change our class structure in 8.1.4. Now my question is what is the best approach to migrate the learning of the adaptive model into the new environment.
**As the class structure has changed a simple migration of the pyADMFactory model will not work.
** uploading responses through the Adaptive Model Management landing page will be very lengthy as we have loads of proposition data.
Do we go with training the models by creating the Data Flows? If so, what is the best way to extract the historical data from the source environment?
Any suggestion/recommendation will be really helpful.
***Edited by Moderator: Lochan to update platform capability tags***
As the class structure has changed a simple migration of the pyADMFactory model will not work. -- I agree with the statement as the ADM factory is the combination of the context and predictors and learning data is the context changes again a new model is created.
Do we go with training the models by creating the Data Flows? You can efficiently train adaptive models in bulk by using a specialized data flow with a dedicated decision strategy. This method is less time consuming than manually uploading historical data to train adaptive models in complex strategies that can contain hundreds of propositions.
If so, what is the best way to extract the historical data from the source environment?you can train your adaptive models with historical or sample customer responses and improve the initial performance of the model and continue using the model after it is trained initially or you can use the concept of smoothening propensity.
To answer your question you need to import the response to a dataset by browsing from the IH tables and follow the below article to configure your data flow and perform the actions accordingly.
Just note that IH usually is configured to contain the responses/outcomes, but customer data will not be in IH.
In new product versions we'll have a way to automatically extract ADM payload (predictor data) for retraining/build external models/analysis but in older product versions this is something you would need to do yourself by sending the payload to e.g. a Cassandra-backed dataset and pull that in.
Many customers already track the payload in big data structures anyway although this often does not include contextual session information.