With linear models, it is hopeless to learn a good weight vectors for some genres. Why? Because a genre like ‘action’ or ‘drama’ is so complex that it cannot be summarized with a few basic features. On the other hand, very specific genres like ‘horror’ or ‘western’ are much easier to learn because of their frequent discriminative keywords. Of course there are also challenging movies from these genres, but the chance to spot a keyword that makes the classification easier is much higher than for a common genre.
To learn something about the feature correlation, we trained linear SVM models for the minor genres like ‘western’, ‘mystery’ or ‘romance’. Then, we used the cosine similarity of the learned weight vectors to determine the how much a pair of genres have in common.
The most representative genre is probably ‘western’ because except for the ‘adventure’ genre, it has not much in common with any other genres:
Intuitively, it means that there is little overlap regarding the features or that the sign of most features is opposite. In our opinion this makes perfectly sense because there are not much horror western and usually western are not for kids either, nor do they have a fantasy or mystery theme very often. In other words, there are cross-overs, but not very often, or more precisely, at least not in the data set.
Another example is the horror genre that is also distinctive:
Since horror is a much broader theme than western, we can seesome correlations with fantasy and sci-fi. Again this makes sense, since a lot of sci-fi movies have a horror theme and the same is true for fantasy movies.
In a nutshell, the correlations confirm that the disentangling of features with linear models is only possible for very specific genres. For more common genres, which either combine themes from multiple genres or genres that are present in almost every movie, like “drama” a linear separation of features is not possible.