We present graph attention networks (GATs), novel neural network architecturesthat operate on graph-structured data, leveraging masked self-attentional layers toaddress the shortcomings of prior methods based on graph convolutions or theirapproximations