Section 7 (page 12) goes over website fingerprinting and how they're able to gain a significant positive match and low false-positive match rate. They can more easily identify noise from signal than previously thought, we're only looking at hidden services (which number far fewer than the number of public websites), hidden service pages change much less frequently than public ones, and they exclude data points that could raise the false positive rate.
Also note that the classifiers here were even better at identifying specific clients than what hidden services they were accessing. So, not only is your hidden service not safe, neither are your users :-)
A likely use of hidden-service-deanonymisation is finding the server that runs a drug market (silk road-style). Those are overwhelmingly likely to be on their own server.
Another commenter already linked to the tor project's response, which suggests that this kind of website profiling attack is less accurate in the real world than their introduction would suggest.
I wonder how easy this is when a lot of websites share both infrastructure and frameworks. How much uniqueness can you extract from a single blog running wordpress when there hundred similar sites hosted on the same server?