What excites me most about this paper isn’t just that there’s a testable hypothesis about why tree-based algorithms work better with tabular data. It’s that it helps us approach a pedagogy (as in how we teach classes to students) for machine learning.