A Incorporates Random Walk Strategy and Enhanced GNN Model for Link Prediction

The link prediction problem has been extensively studied in graph neural networks. However, there are still many problems to be solved in link prediction. when generating node embeddings, using some methods in unsupervised learning can lead to inefficiency, lack of accuracy, and failure to reflect the structural features of the network, which has a significant impact on the accuracy of later link predictions. Therefore, we incorporate a random walk strategy to generate the initial node embeddings in the model, which improves the richness and quality of the node embeddings. Moreover, we add nonlinear feedforward neural networks to the GNN model to increase the nonlinear modeling capability of node features, as well as better representation and learning capability of the network structure. Experimental results on several benchmark datasets show that the proposed method consistently outperforms previous models.