When to stop showing content on recommendation engines?

Interesting question, but I would say this really depends on the recommendation system itself and the purpose of your goal. Let's think it through:

Indeed we could collect the number of logins and create a relative feature of times ignored until chosen. This allows us also to inspect the amount of time a user ignores a show. On the first look, this makes perfect sense to me. However, this will only work if the user decides each time for a new show, from my own experiance I can say this is not valid for all user :D!

Therefore we would need to include also the probability or the fact of the user choosing a new show for given login. To keep things simple let's say we collect only those logins in which the user is deciding for a new tv-show or movie. Then our new feature would make way more sense and is a bit closer to what we can expect from reality.

But there is more. As far as I know (please correct if I'm wrong), Netflix has a sort of prediction or classification on what is the best picture to display for a certain user on each tv-show. In case the pictures change over time, our login feature from above is less useful. Maybe the image made the trick to convince the user of a tv-show or movie - honestly I don't have a clue how to measure this straight away, except maybe with another feature or collected value if the image was different than the last time.

So far so good. Let's say we collect the following:

1. selected/not selected for each movie or tv-show recommended in a session the user picked a new tv show or movie
2. picture changed for each tv show or movie recommended after the last time the user picked a newly recommended item.
3. And maybe to improve our picture algorithm also what kind of picture could have had an impact on the user's decision

Sounds like a plan. But the more I think about it the more comes to my mind, that often my wife is watching on the same Netflix account and our tastes are pretty different for tv-shows or movies we watch in absence of the other. Now, one could say: "yes buddy, this is why netflix offers you profiles". But honestly, I'm not into watching tv that much, so I usually don't care about the profile I'm using. Hence I won't be the only one who shares his profile with a person with another preference, this could have serious influence on our recommendation as well. The profile will be mixed and the precision of our recommendation system will probably decrease.

I'm not sure if this is valid for the mass of Netflix users. We don't have any numbers, so I would assume this is valid for more or less than 30% (just guessing). In this case, you could try to track active user behavior. Like on what kind of recommendations does the user stops for how long, on which tv-shows is he interested in how many seasons it has, how often did the user hover over a movie or show?

The more we think about it the more edge cases and conditions come to my mind. Since I personally don't have a lot of experience with such advanced recommendation systems I would assume that from a technical point of view it would make sense to split the system into two parts. One which gives us a baseline of a larger amount of predictions and one which adjusts the recommendation live, while the user is behaving on the open application to adapt current data and to decide which profile is relevant and applicable for the real person behind the user.

This is really interesting. I'm curious what others (probably with more experience) thing about these questions and about what I provided as an idea.