Do we need all the neurons we utilize? An exploration into costs and scalability for large-scale multi-relational graph learning