An Investigation of the Impact of Normalization Schemes on GCN Modelling

In the modelling of graph convolutional networks (GCNs), typically based on the adjacency matrix of the graph, most studies opt for the symmetric normalized Laplacian as the normalization method for the adjacency matrix. However, there has been little research discussing the impact of alternative normalization methods on the performance of GCN deep learning tasks. Therefore, this paper focuses on the performance of two normalization approaches (symmetric normalized Laplacian and random walk normalized Laplacian) in GCN's node classification task. Additionally, to effectively control the scale and parameters of the network, this study combines a GCN sparsification scheme to draw conclusions. Experimental results on three benchmark graph network datasets indicate that the symmetric normalized Laplacian generally achieves better performance in most cases. However, the results also depend on the selection of sparsification methods and the setting of hyperparameters.