With delayed learning in place our adaptive models are feeding with one negative or positive response once timeout happens. It is fine for most of the channels, however for email it doesn't look complete based on scenarios we have:
We have multiple email content where there is no clickable url to accept/like the offer, in those cases considering 'Open' as positive response (both open and clicks are defined as positive response in adaptive model rule) is enough but the same is a challenge to solve for an email with clickable links/buttons/urls. Because a customer can open an email and do not click it or open and click - in both cases delayed learning will output a positive response, however we can understand it's not the same.
Since we can't give a weightage to any response while feeding response to adaptive, we thought of having two adaptive model rule (one with only open as positive and one with only click as positive) and use them together with a weighting in prioritization formula. However challenge is how to trick delayed learning feature to output both click and open at the end of wait time window?
Or do you have a better solution to avoid the above situation and solve the problem?
***Edited by Moderator Marissa to add Capability tags***
@Gaurav_Kumar I would agree with not mixing two different outcomes in one model. If some models are based on opens and other on clicks, the propensities are not really comparable. That would actually suggest to use only open to prioritize across actions, as only that is available to all your e-mail offers.
You could consider to use the clicks in another way, e.g. upweighting the priorities (in the value calculation). You could have a separate model that predicts clicks (possibly click-given-open) and use that to boost the value in your p*V prioritization. But there is no hard and fast rule how to do this, just like it is not trivial to choose the weights when you would combine two models.