Link Prediction Based on Contrastive Multiple Heterogeneous Graph Convolutional Networks

Link prediction research based on graph neural networks has produced many research results, the technical idea is to use the topology of the graph and the attribute information of the nodes for node representation learning, mapping the graph to a low-dimensional vector space, and then use the link prediction function to measure the similarity of the node representations to determine the likelihood of the existence of links. However, for the link prediction task of multiple heterogeneous attributed networks (AMHEN), Most models overlook the diversity and importance of different relationships between nodes, as well as the impact of the nature of link prediction on the model. In this paper, we propose a link prediction based on Contrastive Multiple Heterogeneous Graph Convolutional Networks (PL-MHGCN) for the problem of link prediction in heterogeneous networks based on graph neural networks. The PL-MHGCN algorithm decouples the multi-network into multiple sub-networks, dynamically learns the edge weights of the sub-networks, and uses a simplified graph convolution to automatically capture heterogeneous metapath information. The algorithm is extended for negative samples, and the model is trained using the agent AUC objective function, which reduces the impact of network sparsity on the link prediction task. The experimental results show that PL-MHGCN is essentially optimal compared to the baseline algorithm and achieves more than 98% prediction accuracy on the Alibaba dataset, and at least 5% improvement compared to the optimal baseline algorithm on IMDB and DBLP.