A very practical and useful collection of hints. I particularly support the point about tailoring the set of informative features. In the end, there is no free lunch for the "core" of the machine learning method (cf., http://ti.arc.nasa.gov/m/profile/dhw/papers/78.pdf).

Now deep learning seems to debunk the more-or-less usefulness of the feature extraction step.

I still don't understand why papers are not dated.