Treeago: Tree-structure aggregation and optimization for graph neural network

Abstract

Graph Convolutional Networks (GCNs) have achieved much success in various graph learning tasks. However, as the number of layers increases, the smoothing of GCNs will over-mix the neighbors’ information, leading output towards space with low expressivity. It is known as the over-smoothing issue. Although several works have refined deep GCNs by optimizing network structure, receptive field, and topology, the over-smoothing issue cannot be completely avoided. In this paper, we propose a recurrent neural network framework for learning graph representation while avoiding over-smoothing effectively, which is the tree-structure aggregation and optimization framework named Treeago. Treeago firstly transforms the irregularly distributed graph into sequential trees. Then, Treeago adopts Tree-LSTM with attention to aggregate important neighbors’ feature information to the graph representation. Tree-LSTM with attention can prevent the mixing of noise neighbors’ information to avoid the over-smoothing issue. Finally, Treeago uses an edge pruning optimization framework based on reinforcement learning to enhance the model’s performance further. Experimental results on multiple real-world datasets show that Treeago effectively avoids over-smoothing and yields state-of-the-art results.