Graph Convolutional Network with Learnable Message Propagation Mechanism

In recent years, graph neural network has become the main paradigm for solving graph analysis tasks, which can easily process high-dimensional data and has a powerful fitting capability. Recent works on graph neural networks have successfully transferred the convolution network in computer vision to graph. Graph convolution network (GCN) has become the classical network framework in graph neural networks due to its simple aggregation approach and favorable theoretical support. When the original graph data are constructed, an adjacency matrix is used to represent the topology, where 0 or 1 indicates whether there is a connection between nodes. Moreover, GCN aggregates node attributes only depends on adjacency matrix. Although it can learn a mapping function, its message propagation mechanism is fixed for a given adjacency matrix. However, for specific downstream tasks, we expect to propagate messages relevant to the downstream task, while a fixed aggregation mode cannot handle this. To this end, we propose a graph convolution neural network with a learnable message propagation mechanism. The original adjacency matrix is adjusted through a learnable weight, so that the message propagation mechanism better adapts to downstream tasks. Experimental results show that the proposed model achieves significant performance in node classification.