In order to get a good prediction of "should watch / shouldn't watch", the system used scores on a 0-10 scale for the amount of 20+ categories present in each film... then each user would give their category preferences on a -5...+5 scale... and the sum of a film's category scores × user preferences, would end up being highly correlated to the user's like/dislike of the film.

From the end user's perspective, it only required entering 20+ preferences... but scoring each film on 20+ categories, proved much more difficult. People would give different scores for their perceived amount of a category in a film, and while the personal sum[score×preference] was highly correlated to their like/dislike verdict, the sum[avg(scores)×preference] was all over the place, and we weren't able to find a way to assign film category scores that would work reasonably well for everyone.

Turns out people not only have different category preferences, but also different category perceptions for the same film.

Maybe revisiting the idea today, with the help of some AI, could find some effective grouping or a different predictor, but back then we just mothballed the whole thing.