We recently came across a business requirement related to action/treatment distribution in a NBA setup.
Below is a summary:
Product: Pega/Marketing 8.4 and NBA Designer is being used for setup.
Adaptive Model - Yes (OOTB Email Click Through Rate updated with our predictors).
Adaptive Model Delayed Learning - is in scope (Offer emails would be sent from a different product and customer response to emails would be delivered to Pega in a file once a day).
Outbound Run with a daily segment refresh option.
Total 10 actions are present under one Issue/Group.
Since we are using NBA, the actions would be going through the OOTB Email_Click_Through Rate adaptive model during a outbound run.
During the 1st two weeks of go-live, business don't want the Adaptive Model to pick the top offer but expecting evenly distribution of 10 actions among eligible customers. so that all 10 actions would be getting equal opportunity for receiving customer response
During those two weeks, Pega would be receiving customer responses to the emails offers being sent daily in a file. we are planning to load the same to Email channel adaptive model using a dataflow/OOTB analytics data sets.
After two weeks into go-live, there would be enough responses collected/fed into the model, then model can start picking the best offer for all customers based on the model learnings.
What is the best approach to implement above requirement? what changes do we make at Action/NBA strategies level to achieve this? we would revert back to initial NBA setup after two weeks.
P.S. Our business is not ready to let model pick best offer from day 1 itself. We had multiple discussions on this topic.
This is one of the common usecase in ADM implementation. I would like to recommend the following two options:
1. Enable ADM from Day1 and start learning from all the customer responses. Train the ADM model from the customer feedback. However, don't use the ADM propensity for offer/action section in your prioritization/arbitration logic until you see the ADM model performance is above 65-70 OR ADM collected enough evidences to predict the positive outcomes.
2. PEGA has solution (smoothen propensity) formula to address the ADM cold-start problem. You can leverage the same to learn/train the ADM models and also making recommendations based on ADM propensity.
Posted: 7 months ago
Updated: 7 months ago
Posted: 2 Nov 2020 6:34 EST Updated: 7 Nov 2020 21:48 EST