We study the structure and evolution of educational assortative mating in the U.S. using 1962-2025 Current Population Survey data. From observed marital patterns, we recover the full matrix of marital gains across educational groups within the equilibrium matching framework of Choo and Siow (2006) and provide a novel characterization of its complementarity structure using Singular Value Decomposition. This approach allows us to infer the number of relevant sorting dimensions, to describe how different educational attainments are valued along each dimension, and to develop aggregate measures of assortative mating based on local odds ratios. We show that the first dimension explains more than 90% of total sorting in all periods and ranks individuals monotonically by educational attainment, consistent with the presence of human capital complementarities à la Becker. On the other hand, higher-order dimensions capture homogamous preferences that help explain the prevalence of same-education couples in the data. We document evidence of a long-run increase in educational assortative mating, with a marked slowdown since the 1990s, although the magnitude of the long-run increase is sensitive to how educational categories are defined, partly due to a coding change in the CPS educational variable in 1992. We also show that high-school dropouts have become increasingly isolated, college and postgraduate groups have converged, and gender asymmetries in returns from sorting have declined.