The other replies are interpreting this as the algorithm failing, but I have interpreted these sorts of things as intentional design choices, wherein they want the recommender to keep trying to diversify your interests so it's harder for you to just quit the service and move to another one which might not have the same variety (or where you'd have to try to "teach" the recommender again). They've determined that the potential benefit is much better than somewhat annoying you.
This interpretation of their behavior is why I've stuck to buying my music (fortunately that's still common for the genres I'm into).
In this instance - and others in my experience listening to the recommendations on this and similar services - "diversify" is used when "dilute" would be more appropriate.
It’s not really though - if it was known in advance that you did not like a track then there would be no reason to recommend it. It’s the classic precision vs recall trade off: I can create a recommendation algorithm that only recommends your favourite song, forever, and that will have perfect precision but miserable recall. To increase recall we have to accept a drop in precision.
I’ve found that most algorithms tend to reinforce a taste by trying to provide more of the same. They rarely try to bring in something that diverges from the pattern. Of course, the libraries have limits and the algorithms will often match against characteristics that you do not consider relevant.
I would love diversifying, but it does not do that. It does "revert to mainstream" trying to push you toward the most generic thing accestible by association from what you like.
This interpretation of their behavior is why I've stuck to buying my music (fortunately that's still common for the genres I'm into).