Abstract: A globally-augmented self-attentive marked temporal point process (GASAMTPP) model is proposed. The model encodes an entity’s static features into a global latent variable to achieve implicit clustering; it applies a self-attention mechanism to encode the historical event sequence that has been concatenated with the global latent variable to capture complex dependencies among the entity’s historical events. Leveraging the resulting comprehensive representation, the conditional intensity function and the mark distribution are parameterized, enabling joint prediction of the next event time and its mark. Experiments on real datasets show that, for next-event time and mark prediction tasks, GASAMTPP outperforms existing mainstream models.