- Abstract: Neural networks are increasingly used as a general purpose approach to learning algorithms over graph structured data. However, techniques for representing graphs as real-valued vectors are still in their infancy. Recent works have proposed several approaches (e.g., graph convolutional networks), but as we show in this paper, these methods have difficulty generalizing to large graphs. In this paper we propose Graph2Seq, an embedding framework that represents graphs as an infinite time-series. By not limiting the representation to a fixed dimension, Graph2Seq naturally scales to graphs of arbitrary size. Moreover, through analysis of a formal computational model we show that an unbounded sequence is necessary for scalability. Graph2Seq is also reversible, allowing full recovery of the graph structure from the sequence. Experimental evaluations of Graph2Seq on a variety of combinatorial optimization problems show strong generalization and strict improvement over state of the art.