A novel link prediction algorithm based on inductive matrix completion

Abstract Link prediction refers to predicting the connection probability between two nodes in terms of existing observable network information, such as network structural topology and node properties. Although traditional similarity-based methods are simple and efficient, their generalization performance varies widely in different networks. In this paper, we propose a novel link prediction approach ICP based on inductive matrix completion, which recoveries node connection probability matrix by applying node features to a low-rank matrix. The approach first explores a comprehensive node feature representation by combining different structural topology information with node importance properties via feature construction and selection. The selected node features are then used as the input of a supervised learning task for solving the low-rank matrix. The node connection probability matrix is finally recovered by a bi-linear function, which predicts the connection probability between two nodes with their features and the low-rank matrix. In order to demonstrate the ICP superiority, we took eleven related efforts including two recent methods proposed in 2020 as baseline methods, and it is shown that ICP has stable performance and good universality in twelve different real networks. Compared with the baseline methods, the improvements of ICP in terms of the average AUC results are ranging from 3.81% ∼ 12.77% and its AUC performance is improved by 0.08% ∼ 3.54% compared with the best baseline method. The limitation of ICP lies in its high computational complexity due to the feature construction, but the complexity can be reduced by replacing complex features with node semantic attributes if there are additional data available. Moreover, it provides a potential link prediction solution for large-scale networks, since inductive matrix completion is a supervised learning task, in which the underlying low-rank matrix can be solved by representative nodes instead of all their nodes.