Now that we created scikit-learn-contrib, one of scikit-learn's roles is to ease the creation of new contrib projects. So we really need to sort out the public vs. private utils issue once and for all. Despite many discussions, AFAIK, we haven't reached a consensus or come up with a realistic plan.

Whether we like it or not, people do use utils and they complain when utils break.

Based on this I propose that

1. All current utils are now officially public and if we change them, we need to apply the usual 2 release compatibility rule.
2. From now on, new utils should be made private (i.e., start with a leading underscore) unless we have a good reason to make them public. In other words, we should think twice before making an util public.