Learning a Canonical Basis of Human Preferences from Binary Ratings