SINCERE: Sequential Interaction Networks representation learning on Co-Evolving RiEmannian manifolds