We have a requirement to create a match between ADM propensity and business value of a proposition. In some of the cases, business value is high for few products but ADM recommendation is less. Based on some analysis, I found four different approaches. I am trying to choose the best one out of four.
Approach 1: Multiply the propensity * Business value before sending the offer/action to the channel.
Approach 2: Use a classification matrix and derive a value for business classification like High/Medium/Low and pass this value to ADM model as predictor.
Approach 3: Only pass the high business value propositions to ADM. Filter our the ones which are below the threshold.
Approach 4: Reduce the final propensity(.FinalPropensity = .pyPropensity/n times) if the business value is high and increase the final propensity(.FinalPropensity = .pyPropensity * n times) if business value is low. If both are matching each other medium business value with medium propensity, please use the same propensity (.FinalPropensity = .pyPropensity) as final propensity.
What could be best approach to create a match between ADM propensity and business value? Any suggestions. Thanks in advance.
Note: This should not impact ADM learning/performance.
***Edited by Moderator Marissa to update Platform Capability tags****
The standard approach is to let the models do what they do best - predict propensities - and use value (or any levers) in the prioritization of the actions. When using only the propensities for prioritization you will be offering things that customers are most likely to accept (very empathetic) but typically you want to factor in some notion of value so p*V would be the usual setup, and this is built into the NBA Designer framework.
(PS are you using the NBAD framework?)
(PS2 indeed this approach does not affect how the models learn - they still learn from what gets accepted, regardless of how it happens to be prioritized)
Posted: 1 month ago
Posted: 10 May 2021 6:05 EDT
Guru Deshkulkarni (Guru Deshkulkarni)
Senior Manager, Application Enablement
@Nizam For more information on balancing propensity and value, please go through the Pega Decisioning Consultant mission on Pega Academy. In Cross-sell on the Web mission, Arbitrating between actions module, you'll see how to balance customer relevance and business priorities using the formula P * C * V * L. Customer relevance is represented by the propensity from the adaptive models (P) and the customer weighting (C). Business priority is represented by action value (V) and business levers (L). In the Next-Best-Action Designer you configure how this formula P * C * V * L plays a role during arbitration.
At Vodafone, we've been using adaptive models for most of our journeys. This is what we decided as best;
SmoothedPropensity * Price * OfferValue
SmoothedPropensity is the formula from the pega training basically.
Price helps us meet the business needs if the propensity of the 2 offers close, but one is the more expensive business wants to offer that one. By using price in the calculation, we are trying to maximize uplift and take rate. As you know, adaptive models are already trying to maximize success rate(take rate), so having a price is helping to maximize take rate.
Offer Value; we are managing this at the proposition side; each offer has a property normally set to 1; if the business needs to increase the frequency of an offer, they are basically increasing the offer value from 1 to 1.2.