DEEP GRAPH TREE NETWORKS

Nan Wu, Chaofan Wang

Keywords: graph tree networks, graph tree convolution networks, graph tree attention networks, GNNs

Abstract: We propose Graph Tree Networks (GTree), a self-interpretive deep graph neural network architecture which originates from the tree representation of the graphs. In the tree representation, each node forms its own tree where the node itself is the root node and all its neighbors up to hop-k are the subnodes. Under the tree representation, the message propagates upward from the leaf nodes to the root node naturally and straightforwardly to update the root node's hidden features. This message passing (or neighborhood aggregation) scheme is essentially different from that in the vanilla GCN, GAT and many of their derivatives, and is demonstrated experimentally a superior message passing scheme. Models adopting this scheme has the capability of going deep. Two scalable graph learning models are proposed within this GTree network architecture - Graph Tree Convolution Network (GTCN) and Graph Tree Attention Network (GTAN), with demonstrated state-of-the-art performances on several benchmark datasets. The deep capability is also demonstrated for both models.

One-sentence Summary: We propose a self-interpretive deep GNN architecture: Graph Tree Networks, and two models within this architecture: Graph Tree Convolution Networks and Graph Tree Attention Networks with demonstrated state-of-the-art performances.