This paper addresses the problem of learning a sequence model for marked temporal point processes. The problem has been studied in the literature and is of interest to the community. However, the paper is not written clearly and lacks coherence, there are many typos, and the authors use a language that is not appropriate for a scientific paper. One of my criticisms of the paper is the use of "causal-graphs" in the paper. The term causal generally refers to causal statements which answer counter-factual statements of the form "what if ...", the authors use this term extensively to describe the graph which they recover using their method, while the recovered graph may not be the causal graph in the sense often understood by the community. My other criticism of the paper is lack of coherence between sections of the paper. This makes understanding the motivation behind every section very difficult. In its current form, it is very difficult to follow the logic behind each step and hence understanding the paper.

The paper presents several fresh ideas and techniques for leanring high-dimensional event sequence in the continuous time space, which is challenging and still remains open. Specifically, the authors devise a random walk based approach among the local causal graph of markers, for high-dimensional event generation with a linear complexity efficient implementation, after a rigrous theoretical study. Meanwhile, a latent intensity model is devised without explicitly defining the network structure. This model can be also useful for high-dimensional marker space, as it instead model the latent space rather than the raw marker space which can be intractable. The distance between two high-dimensional event sequences is modeled via adversarial imitation learning, which makes the distance computing more tractable and flexible. Overall the paper is well written and the technical novelty is solid. The experiments are comprehensive.

This paper targets modeling the temporal dynamics in high-dimension marked event sequences without any given causal network of markers. This is a difficult problem and rarely studies in previous studies. The authors propose a seminal adversarial imitation learning framework which consists of a generator model that takes a bottom-up view to imitate the generation process of event sequences and an interpretable and efficient random walk based sampling approach to generate the next event. The methods are novel and have good intuition. The paper is well written and easy to follow. One suggestion from me is not to put too many things in a single section ``Introduction and Related Works''. The experiments are well conducted and analyzed. Some conclusions are demonstrated. But I have some concerns about experiments part. Check ``improvements'' part for details.