Does anyone know at a high level how pega works out the propensity for a proposition whose model is yet to be trained for the first time? i.e. The minim number of responses before the model is updated has not been reached.
Is is limited to basic principles like acceptance rate from the responses received so far?
I noticed that model evidence builds up although only a fraction of the actual responses are used before the model is updated. How does Pega decide which responses contribute to the propensity calculation and is it actually processing these responses and evaluating predictors etc or again is it based on more basic principles at this stage?
Any light you can shed is much appreciated. We use the smooth propensity concept in our implementation and as props converge towards actual propensity it's important to understand the make up of this propensity given that Pega cannot discriminate between customer profiles before a model has been through at least one round of training.
***Edited by Moderator Marissa to update Platform Capability tags****
In the initial starting phase of a model instance, before the model is trained and uses active predictors, the propensity will be based on the evidence received. It won't use any predictors but will just be based on #positives and #negatives (=base propensity).
The update frequency of models is defined in the adaptive model rule settings and will be triggered at least every time N new responses are received. But initial learning cycles already kick in when 50 responses have been received with next cycles at 100 and 500 responses. Once the model has active predictors the propensity returned is based on the active predictors.
P.S. a learning cycle may not yet result in a model with active predictors, this can for example happen when the number of positives received is still too low