I was involved with design and planning for a solution to predict unscheduled maintenance and warranty claims for vehicles. I think that we had a pretty similar

I was involved with design and planning for a solution to predict unscheduled maintenance and warranty claims for vehicles. I think that we had a pretty similar problem. Unfortunately I did not stay with the company I was working for, so I can't speak to the effectiveness of the approach. A simplified example is below. There is tons of scheduled maintenance. And there are also lots of common unscheduled maintenance items. These were basically worthless in any naive prediction. Basically "Ah, I see that an engine failure is always preceded by dozens of oil changes." There's a tiny predictive value, because number of oil changes is a proxy for age/use; older engines are more likely to fail. But that sort of prediction wasn't helpful. We ended up proposing a strategy that I shorthand to "without which, not". Basically we wanted to classify the events which precede a specific type of failure, but which we don't see in similar vehicles without the failure. This is grossly oversimplified. An analog would be the idea of necessary vs sufficient pre-conditions. We were trying to identify the sufficient pre-conditions. To draw the analogy to a recommendation engine, a common failure mode is that popular things get recommended a lot. Another failure mode is that they basically become genre filters. This is discussed elsewhere. So in a movie context, we wouldn't be looking for "what do other people who like this movie watch"? Because that's a genre popularity contest. We would ask "Given that a user likes Movie A, what did they watch before that other Movie A likers didn't watch?" And then look for clusters there. Again, I'm grossly oversimplifying this, because I didn't get to implement the solution I mentioned. The approach seemed promising, though. You might consider this a nuance on "down-weighting popular things", but it's not quite that.