Empowering Graph Representation Learning with Paired Training and Graph Co-Attention

Through many recent advances in graph representation learning, performance achieved on tasks involving graph-structured data has substantially increased in recent years---mostly on tasks involving node-level predictions. The setup of prediction tasks over entire graphs (such as property prediction for a molecule, or side-effect prediction for a drug), however, proves to be more challenging, as the algorithm must combine evidence about several structurally relevant patches of the graph into a single prediction. Most prior work attempts to predict these graph-level properties while considering only one graph at a time---not allowing the learner to directly leverage structural similarities and motifs across graphs. Here we propose a setup in which a graph neural network receives pairs of graphs at once, and extend it with a co-attentional layer that allows node representations to easily exchange structural information across them. We first show that such a setup provides natural benefits on a pairwise graph classification task (drug-drug interaction prediction), and then expand to a more generic graph regression setup: enhancing predictions over QM9, a standard molecular prediction benchmark. Our setup is flexible, powerful and makes no assumptions about the underlying dataset properties, beyond anticipating the existence of multiple training graphs.