TreeRNN: Topology-Preserving Deep GraphEmbedding and Learning

In contrast to the literature where the graph local patterns are captured by customized graph kernels, in this paper we study the problem of how to effectively and efficiently transfer such graphs into image space so that order-sensitive networks such as recurrent neural networks (RNNs) can better extract local pattern in this regularized forms. To this end, we propose a novel topology-preserving graph embedding scheme that transfers the graphs into image space via a graph-tree-image projection, which explicitly present the order of graph nodes on the corresponding graph-trees. Addition to the projection, we propose TreeRNN, a 2D RNN architecture that recurrently integrates the graph nodes along with rows and columns of the graph-tree-images to help classify the graphs. At last, we manage to demonstrate a comparable performance on graph classification datasets including MUTAG, PTC, and NCI1.