Implementation of Node Classification Algorithm Based on Graph Neural Network

With the research and development of Graph Neural Networks (GNNs), GNN has shown very good results in link prediction, node classification, social network and other applications. In this paper, the node classification algorithm based on GNN is implemented by software, and the neural network models that need hardware acceleration are selected and trained. The comparative experiments are conducted on Cora, CiteSeer and PubMed citation network datasets respectively. Through the model training of the combination of different aggregation update functions, the comprehensive analysis of the experimental results shows that the combination of message passing layer functions used in this paper has the best effect, and the test accuracy in three data sets reaches 77%, 59% and 75% respectively. In order to better deploy the network model on the hardware, the symmetric quantization operation is carried out to reduce the parameters, so as to achieve the acceleration of the software part. The experimental results show that the accuracy of the quantized model is almost unchanged.