Of all the music recommendation systems out there, this one is my favorite:
It has an uber simple interface (you don't even need to log in) but no other system gets me as well as this one.
What surprises me in the given article is the graph on the last page. Why was the treatment group still listening more weeks after they already were getting the old recommendations again?
I wish the graph would reach out further into the past then just one week before the test. So one could get a feel for how much of the variance in the graph is random.