Representation learning in multiplex graphs: Where and how to fuse information?