Graph Convolutional Networks with Adaptive Frequency and Dynamic Node Embedding

Over-smoothing is the core bottleneck of deep neural network construction. In order to alleviate the over-smoothing phenomenon in the process of graph learning, we propose a graph convolution network based on adaptive frequency and dynamic node embedding (GCNFN). On the one hand, high-frequency and low-frequency information can help the learning of graph convolution network, so we introduce an adaptive mechanism to improve the aggregation function, so that it can adaptively aggregate information of different frequencies. On the other hand, due to the difference of the suitable receiving domain of nodes, that is, the range of information that nodes can aggregate to other nodes, the inappropriate receiving domain will result in insufficient discrimination of node embedding. Therefore, we use the dynamic node embedding mechanism to adjust and fuse the output of each layer of the network, balance the node information of the whole local area and the local area, and generate more discriminative node embedding to alleviate the problem of excessive smoothing. Experiments on four public datasets show that GCNFN model can achieve better learning accuracy than the comparison model, and maintain higher classification accuracy when appropriately increasing the number of network layers.