Improving Graph Neural Network Models in Link Prediction Task via A Policy-Based Training Method

Abstract

Graph neural network (GNN), as a widely used deep learning model in processing graph-structured data, has attracted numerous studies to apply it in the link prediction task. In these studies, observed edges in a network are utilized as positive samples, and unobserved edges are randomly sampled as negative ones. However, there are problems in randomly sampling unobserved edges as negative samples. First, some unobserved edges are missing edges that are existing edges in the network. Second, some unobserved edges can be easily distinguished from the observed edges, which cannot contribute sufficiently to the prediction task. Therefore, using the randomly sampled unobserved edges directly as negative samples is difficult to make GNN models achieve satisfactory prediction performance in the link prediction task. To address this issue, we propose a policy-based training method (PbTRM) to improve the quality of negative samples. In the proposed PbTRM, a negative sample selector generates the state vectors of the randomly sampled unobserved edges and determines whether to select them as negative samples. We perform experiments with three GNN models on two standard datasets. The results show that the proposed PbTRM can enhance the prediction performance of GNN models in the link prediction task.