Adaptive Propagation Graph Convolutional Networks Based on Attention Mechanism

The main steps in a graph neural network are message propagation and aggregation between nodes. Message propagation allows messages from distant nodes in the graph to be transmitted to the central node, while feature aggregation allows the central node to obtain messages regarding its neighbors and update itself, so that it can express deep-layer features. Because the graph structure data have no local translation invariance, the number of neighbors of each central node is different, and there is no order, there are two difficulties: (1) how to design a reliable message propagation method to better express all network topologies; (2) how to design a feature aggregation function so that it can weigh local features and global features. In this paper, a new adaptive propagation graph convolutional network model based on the attention mechanism (APAT-GCN) is proposed, which enables GNNs to adaptively complete the process of message propagation and feature aggregation, according to the neighbors of the central node, and set the influence degree of local and global messages on the aggregation of the central node. Compared with other classical models, this method is superior to the baseline model and can improve the accuracy of node- and graph-level classification tasks in downstream tasks.