The Pt/CeO2 system plays a central role in heterogeneous catalysis, with research focusing on two directions: Pt single atoms for maximum metal utilization and Pt clusters for greater stability. However, the dynamic restructuring of Pt on CeO2 remains poorly understood due to the computational complexity of cerium. Here we investigate this process using density functional theory (DFT) combined with machine learning interatomic potentials (MLIPs). Our simulations reveal facet-dependent aggregation of Pt single atoms into clusters at elevated temperatures. Pt atoms exhibit higher mobility on the CeO2(111) surface, which promotes clustering, particularly near the boundaries between (111) terraces and the (110)/(100) facets. Reactivity analysis shows that methane activation has high barriers on isolated Pt atoms, whereas in situ formed Pt clusters significantly lower the activation barrier. Notably, clusters formed through single-atom aggregation exhibit higher catalytic activity than directly adsorbed clusters due to optimized coordination environments and stronger electronic coupling developed during restructuring. These results provide atomistic insights into Pt/CeO2 interfaces and highlight the potential of MLIPs for large-scale simulations of catalytic systems with complex electronic structures.