Neural message passing on high order paths

Abstract Graph neural networks have achieved impressive results in predicting molecular properties, but they do not directly account for local and hidden structures the graph such as functional groups geometry. At each propagation step, aggregate only over first order neighbours can learn about important information contained subsequent well relationships between those higher connections—over many steps. In this work, we generalize nets to pass messages across paths. This allows propagate various levels substructures of graph. We demonstrate our model on a few tasks property prediction.