There's a slight weirdness in this paper in the way they measure similarity (which is admittedly a very hard problem). It's entirely derived from the recommenda

There's a slight weirdness in this paper in the way they measure similarity (which is admittedly a very hard problem). It's entirely derived from the recommendation networks of IMDB and Amazon, which are mostly built on the basis of "people who liked/watched this also liked/watched these movies". Presumably this should measure similarity on many fronts, taking into account the wide array of each user's subjective factors, including screenwriting style as mentioned in joe_the_user's comment, for example. I think it's a pretty wobbly base for the research, though. It doesn't make it any less interesting, but even assuming that the recommendation networks perfectly represent the network of user preferences, does this really reflect movie similarity? People stick to directors or actors that they like, even across very dissimilar movies, but this still counts as a form of "similarity" in this paper. And the Amazon network "finds items that customers tend to purchase together", which might not represent similarity at all, but complementarity! Also, consider a film that's extremely similar to another one, but is marketed in such a way as to reach a different audience. This happens when marketers deem that "lying" about the movie in the trailers and marketing campaign will make for better box-office/video sales. Again, this might bend the recommendation networks in a way that diverges from actual movie similarity. Despite this, the study hardly seems to question their measure of similarity at any point (I haven't read everything, though). Or rather, they redefine "similarity" to mean "proximity in the recommendation networks", which includes all these hidden factors that might overlap with other things they're comparing against. As I said, though: it's a hard problem.