I mean, if you read about how current industry-standard recommendation systems work, this is pretty bang on, I think? (I am not a data scientist/ML person, as a disclaimer.)
If e.g. retention correlates to watch time (or some other metric like "diversity of content enageged with"), then you will optimize for the short list of metrics that show high correlation. The incentive to have a top-tier experience that gets the customer what they want and then back off the platform is not aligned with the goal of maintaining subscription revenue.
You want them to watch the next thing, not the best thing.
If e.g. retention correlates to watch time (or some other metric like "diversity of content enageged with"), then you will optimize for the short list of metrics that show high correlation. The incentive to have a top-tier experience that gets the customer what they want and then back off the platform is not aligned with the goal of maintaining subscription revenue.
You want them to watch the next thing, not the best thing.