“…In recent years, neural marked temporal point processes (MTPP) have shown a significant promise in modeling a variety of continuoustime sequences in healthcare [33,34], finance [3,47], education [43], and social networks [29,48,49]. However, standard MTPP have a limited modeling ability for CTAS as: (i) they assume a homogeneity among sequences, i.e., they cannot distinguish between two sequences of similar actions but with different time duration; (ii) in a CTAS, an action may finish before the start of the next action and thus, to model this empty time interval an MTPP must introduce a new action type, i.e., NULL or end-action which may lead to an unwarranted increase in the types of actions to be modeled; and (iii) they cannot encapsulate the additional features associated with an action, for e.g., minimum time for completion, necessary previous actions, or can be extended to sequence generation.…”