Iteration Graph Network

Graph neural networks are widespreadly used in the field of graph data analysis and processing. Recent methods either reduce the spatial receptive field for low algorithm complexity, or greatly lose efficiency in order to realize attention mechanism. To tackle this issue, we propose Iteration Graph Network (IGN), which uses an iterative inversion method to aggregate node feature and the k-localized neighbor information of nodes. In the graph-based semi-supervised node classification task, our method surpasses the state-of-the-art method in the benchmark datasets and experiment conclusion show that our model outperforms graph attention networks (GAT) and is more than 3 times faster than graph attention networks, consumes more than 6 times less memory than GAT. Our code will be make publicly available.