Ceria-based single-atom catalysts present complex electronic structures due to the dynamic electron transfer between the metal atoms and the semiconductor oxide support. Understanding these materials implies retrieving all states in these electronic ensembles, which can be limiting if done via density functional theory. Here, we propose a data-driven approach to obtain a parsimonious model identifying the appearance of dynamic charge transfer for the single atoms (SAs). We first constructed a database of (701) electronic configurations for the group 9–11 metals on CeO₂(100). Feature Selection based on predictive Elastic Net and Random Forest models highlights eight fundamental variables: atomic number, ionization potential, size, and metal coordination, metal–oxygen bond strengths, surface strain, and Coulomb interactions. With these variables a Bayesian algorithm yields an expression for the adsorption energies of SAs in ground and low-lying excited states. Our work paves the way towards understanding electronic structure complexity in metal/oxide interfaces.