Abstract Mesoscale eddies, ubiquitous in the global ocean, play a key role in the climate system by stirring and mixing key tracers. Estimating, understanding, and predicting eddy diffusivity is of great significance for designing suitable eddy parameterization schemes for coarse-resolution climate models. This is because climate model results are sensitive to the choice of eddy diffusivity magnitudes. Using 24-yr satellite altimeter data and a Lagrangian approach, we estimate time-dependent global surface cross-stream eddy diffusivities. We found that eddy diffusivity has nonnegligible temporal variability, and the regionally averaged eddy diffusivity is significantly correlated with the climate indices, including the North Pacific Gyre Oscillation, Atlantic multidecadal oscillation, El Niño–Southern Oscillation, Pacific decadal oscillation, and dipole mode index. We also found that, compared to the suppressed mixing length theory, random forest (RF) is more effective in capturing the temporal variability of regionally averaged eddy diffusivity. Our results indicate the need for using time-dependent eddy mixing coefficients in climate models and demonstrate the advantage of RF in predicting mixing temporal variability. Significance Statement Mixing induced by ocean eddies can greatly modulate the ocean circulation and climate variability. Steady eddy mixing coefficients are often specified in coarse-resolution climate models. However, using satellite observations, we show that the eddy mixing rate has significant temporal variability at the global ocean surface. The regional temporal variability of eddy mixing is linked with large-scale climate variability (e.g., North Pacific Gyre Oscillation and Atlantic multidecadal oscillation). We found that random forest, a user-friendly machine learning algorithm, is a better tool to predict the mixing temporal variability than the conventional mixing theory. This study suggests the possibility of improving climate model performance by using time-dependent eddy mixing coefficients inferred from machine learning methods.