Recently, there has been a surge of interest in learning representation of graphstructured data that are dynamically evolving. However, current dynamic graph learning methods lack a principled way in modeling temporal, multi-relational, and concurrent interactions between nodes—a limitation that is