Learning Structure-enhanced Temporal Point Processes with Gromov-Wasserstein Regularization