Cation ordering in multication perovskites is related to many important material properties and performances, but computational determination of the cation ordering remains a major challenge. Here, we propose a new computational approach by introducing a machine learning recommender system into the basin-hopping framework (RBH) for optimizing cation ordering. Taking the electrocatalyst Ba0.5Sr0.5Co0.8Fe0.2O3 (BSCF5582) as a showcase example, we found that the efficiency of RBH in identifying low-energy configurations outperforms the methods of cluster expansion and conventional basin-hopping. The RBH results revealed that the BSCF5582 catalyst tended to have a layered ordering of A-site cations and disordered B-site cations both in bulk and on the surfaces. Further, on the A-site-terminated surface, we found the segregation of large Ba atoms. Similarly, on the A-site- terminated surface of the recently developed Cs0.2Sr0.8Co0.4Fe0.6O3 (CSCF2846) catalyst, layered ordering at the A-site and surface enrichment of large Cs atoms were also found. The layered ordering was robust against thermal effects, as found from molecular dynamics simulations at 800 K. This work provides a new approach for thermodynamic global optimization of chemical ordering in multicomponent materials.