Link Value Estimation Based Graph Attention Network for Link Prediction in Complex Networks

Abstract

Link prediction in complex networks aims to discover hidden or forthcoming links between network nodes, and it is widely used in areas such as knowledge graphs. Existing Graph Neural Networks (GNNs) often only consider whether nodes are connected or calculate the weight of the links through node features when they apply network structure and features to learn node representation in the manner of neighborhood aggregation, while neglecting the intrinsic value of links. This paper analyzes the value of links based on network structure and proposes corresponding evaluation metrics. It integrates the value of links into the construction and training of the link prediction graph attention network, not only improving the performance of the link prediction, but also providing theoretical support for the interpretability of the prediction results. Extensive link prediction experiments were carried out on representative open graph benchmark data, and the results show that the link prediction framework proposed in this paper has good performance and generalization capabilities.