Scholarly communication research often relies on comprehensive subject classifications to evaluate research produced within or across disciplines. Such use of classification systems is less related to information retrieval and more aligned with the types of knowledge discovery tasks described by Beghtol (2003) in her discussion of naive classification. In this thesis project in progress, we investigate the machine learning processes used to generate the Microsoft Academic Graph and OpenAlex subject classification systems to better understand how this classification supports knowledge discovery in a research evaluation context, and in what ways it might be made more effective in that context.