NGLinker: Link prediction for node featureless networks

Abstract

Link prediction is a paradigmatic problem with tremendous real-world applications in network science, which aims to infer missing links or future links based on currently observed partial nodes and links. However, conventional link prediction models are based on network structure, with relatively low prediction accuracy and lack universality and scalability. The performance of link prediction based on machine learning and artificial features is greatly influenced by subjective consciousness. Although graph embedding learning (GEL) models can avoid these shortcomings, it still poses some challenges. Because GEL models are generally based on random walks and graph neural networks (GNNs), their prediction accuracy is relatively ineffective, making them unsuitable for revealing hidden information in node featureless networks. To address these challenges, we present NGLinker, a new link prediction model based on Node2vec and GraphSage, which can reconcile the performance and accuracy in a node featureless network. Rather than learning node features with label information, NGLinker depends only on the local network structure. Quantitatively, we observe superior prediction accuracy of NGLinker and lab test imputations compared to the state-of-the-art models, which strongly supports that using NGLinker to predict three public networks and one private network and then conduct prediction results is feasible and effective. The NGLinker can not only achieve prediction accuracy in terms of precision and area under the receiver operating characteristic curve (AUC) but also acquire strong universality and scalability. The NGLinker model enlarges the application of the GNNs to node featureless networks.