While testing Adaptive Models we faced some questions regarding propensity calculation.
According to PEGA's course manual (Student Guide - PEGA_Decisioning), the adaptive model computes a propensity for each proposition input. However, since it is a learn on the fly model and the likelihood of a customer accepts an offer slightly changes if a customer with similar profile accepts the same offer, we were expecting to get a propensity per proposition based on customer properties, i.e., two customers with different properties should have different propensities for the same proposition.
On the model report – Score Distribution tab we have propensities for each bin, however when we call the model from a strategy, the output propensity is always the same.
How is the output propensity calculated?
Shouldn’t it be dependent on the customer?
Note: The Output Propensity (for the model in attachment) is always 0,5620.
***Edited by Moderator: Pooja Gadige to add platform capability tag***
The Adaptive Model does learn per customer profile. However, when there aren't enough variations in the gathered responses, which is typically the case when the model is new and hasn't had the opportunity to gather enough responses, all customer profiles are binned together. Therefore, all customers receive the same propensity.
As the model matures, i.e. gathers enough responses with enough customer profile variations, you'll see more bins in the score distribution, which means different propensity values for different customer profiles. Here's a score distribution of a model that has learned from ~300 responses.