IAGNN:Graph Neural Networks with Information Augmentation

Graph convolutional neural networks use the message passing mechanism to stack layers to expand the perceptual field of view, as a way to expand the rich neighborhood information for the central node, which brings effective benefits for the processing of non-Euclidean graph data. However, the mechanism has some limitations, on the one hand, in small data samples, the multi-layer network of graph neural networks may lead to some node attributes approximating similar fixed values, which leads to the phenomenon of over-smoothing. On the other hand graph neural networks are overly dependent on the node features already present in the graph data, and the trained model has poor adaptability in the data changing environment, i.e., low generalization performance. In order to address the above limitations of graph neural networks, this paper proposes a graph neural network model with information augmentation (IAGNN), which utilizes a conditional variational auto encoder in the pre-training process to focus on local information enrichment of low-degree nodes while avoiding excessive stacking of graph layers. Thereafter, random feature masking enhancement techniques are utilized to mask some of the feature representations of the nodes, thereby enhancing the model generalization capability. It is shown through experiments and analysis on six public datasets that our proposed IAGNN model can achieve good benefits on the downstream task of graph classification.