According to the nextflix prize leaderborad at
http://www.netflixprize.com/leaderboard.html, this result wouldn't even be in the top 1000. More than that, due to the absence of compositional structure in the data I wouldn't expect this approach to work well. Would love to hear more insights, I'm digging into the paper now.
All that said, open code for running this on arbitrary data is great to have! A thanks to Oleksii Kuchaiev.