Recommendation systems for research papers have in my opinion not taken off because they focus on the wrong things. In Google Scholar it is mainly based on what you have published, which is not helpful for people who have not yet published or people in industry who do not publish but may still be interested in research in their field. Another reason is that recommenders such as PubChase focus on biomedical research, which is only relevant to few people. A final reason is that academics, who are likely the main but not sole users, have a different agenda: read all the papers in a particular field. Industry people may benefit more from such a recommendation system because they do not have the time or inclination to check out the latest papers.
Personally, I think a recommendation engine for the articles in your Paperpile library (maybe with a like/dislike button system) would provide sufficient value. I am not an academic researcher, so I do not need to read every single paper in my field, but I would benefit from a system where papers are recommended to me based on my interests/library. I already like the 'Cited by' feature and its easy-to-include-in-my-library feature but something that would use to auto-tag articles (e.g. a simple LDA) and scrape reputable sources for similar articles would be very valuable. Another option would be to have collaborative filtering, i.e. based on people who have the same articles in their library. This would only work well with lots of users. In my opinion a company that did this very well, but unfortunately had to close, was Prismatic, the news aggregator app. By saying what articles were of interest and which ones weren't it found a lot of interesting news that was missed by almost all other apps.