I am also a big fan of the kernel two-sample tests (still the best set of methods for the problem). These seem little known somehow. But they are a family of methods that are incredibly powerful, practical, and easy to use. But without any exposure to RKHS and kernels they now seem very exotic.
