Rooted Subtrees Recursive Neural Networks on Graphs
Fangzhou Liang, Yueming Lu · 2021
Graph classification is a fundamental task via rightly recognizing patterns of graph structure or sub-structures. We propose a novel end-to-end framework, RSRNN, for classifying general node-labeled graphs, which is composed of extracting Rooted Subtrees from graphs and learning Recursive Neural Network on subtrees. For generalizing the existed Tree-LSTMs to graph problems, we develop a new branch attention mechanism for fitting RSRNN to arbitrary rooted subtrees structure. This framework is tested on benchmark data sets and major results prove that it is competitive with the state-of-the-art of a convolutional neural network and graph kernels.