Review #1
This paper studies an important problem in vehicle trajectory prediction and interpretation - understanding the causality underlying vehicle interactions. They also mention the limitation of previous work and the novelty (the first work based on Granger Causality). The proposed method combines GVAR and vehicle/road information, which outperforms the pure GVAR method, as is shown in the experiments. Also, the method will benefit general trajectory prediction and personalization. However, the figures in this paper are not very clear and easy to understand, although there are some explanations in the main context. And it would be good if there is more baseline compared.

Review #2
This paper's authors propose using Granger causality to interpret interactions between different agents (vehicles) over time. Specifically, the proposed algorithm could answer if there are interactions, when the interactions start, and the interaction intensities. They validate the algorithm with the INTERACTION dataset and self-collected field data. In general, I think the research is well-motivated; the authors propose a novel idea of using GC to understand vehicular movements and interaction, the system is well-explained, and the evaluation results on real-world data look promising.
The whole paper is very informative. Nevertheless, I find some important information is missing. First, I suggest the authors further explain Fig. 1. Indeed, as we could notice, Fig.1 basically shows the system overview with many details of the proposed system. I want the authors to refer to this figure when explaining the equations in the following subsections. This can make the section more readable.  Second, I would like the authors to provide more details about the system training in Section 3 (both use cases).  
Another concern I have is that the evaluation only involves limited data. In Section 3, the authors mentioned validating the system using the INTERACTION dataset. However, they only picked one case from the dataset for evaluation. I wonder if the proposed system could apply to different traffic scenarios and maintain decent accuracy. Moreover, the provided results (Fig. 5) look interesting. Nevertheless, as the authors mentioned in the limitation, I wonder about the accuracy of the proposed system since the authors do not have the ground truth.


Review #3
The paper addresses a very relevant problem and presents interesting and novel ideas. More specifically, the paper applies Granger causality to study interaction between drivers. The presentation is clear and the proposed approach is technically sound. 
One minor weakness is that the performance assessment remains somewhat circumstancial, rather than quantitative.
Minor comments: The acronym SENN is not defined.