In the previous setup, samples where the best datastructure was the
current considered (the winning samples, in a way), were tossed away.
I think keeping them helps the classifier notice what works.

This improved the classification accuracy with only software features.