Constructing categories: Moving beyond protected classes in algorithmic fairness