Our client would like to start optimizing adaptive models for conversion events, such as enrolling in certain programs, without losing the ability to optimize for clicks. What is the best way to configure this?
Models are live for both web and email, currently.
If I include both Click and Enrollment as positive outcomes in the same adaptive model rule, how would this affect model learning?
Is it better to create a separate model that only defines Enrollment as a positive outcome, and then figure out some way to use both the Click and Enrollment propensities at decision time?
Another solution I am considering is using an Enrollment-optimized model to identify a next-best-ACTION, and then a Click-optimized model to identify the next best TREATMENT. However, I haven't evaluated the extent to which changes to NBAStrategyFramework and its substrategies would be needed to implement this solution.
Any thoughts you could share on this would be much appreciated.
Clicks are often an excellent proxy for the final conversion, so there's not always a need to track the final conversion (in fact most solutions won't do this). It can be a technical challenge to retrieve that feedback because it could come from another system, for example a fulfilment system and there's usually a longer waiting period involved.
Predicting conversion can be done by only recording conversion responses, ignoring the initial click.
Note: when predicting conversion be careful not to compare the propensity to convert for some offers with propensity to click for other offers as the probability to convert will downscale the propensity compared to the click propensity.
If the effectiveness of the click is to be taken into account, you can use an advanced data science pattern that predicts 'click followed by conversion'. This will only be useful if the profile of customers converting is expected to be very different from the profile of those that click, if not then stick with click as a good proxy.