Graph networks rely on the graph structure existing before hand which is not always the case. Here the authors present a novel method for learning the graph structure (using bilevel programming) simultaneously with training a GCN on the resulting graph.