Highlighting Link Prediction in Bipartite Networks via Structural Perturbation

Abstract

Most of the link prediction algorithms for bipartite networks assume that the generation of link is based on a predefined prior assumption. However, for the real-world bipartite networks, the generation mechanism of link is still ambiguous due to their complexity. Consequently, these methods are not always obtaining satisfactory results in all the cases, as each network has its unique generation mechanisms of link. In this paper, by introducing the structure perturbation theory, we propose a non-parameter bipartite structural perturbation method (BiSPM) to predict unobserved links without making any assumption for link formation mechanism. We first make a small perturbation for the training set of a bipartite network. Then for the perturbed network, we use singular value decomposition to obtain the singular vectors and singular values. Finally, during the process of network reconstruction, we hypothesize that when the topological structure meets tiny disturbance in the bipartite network, the coordinate system (singular vectors) of the projection is invariant, but the scaling factors (singular values) will create a tiny change from the perspective of complex system stability. Thus, a new matrix is constructed for prediction by changing the singular values of the perturbed networks while fixing the singular vectors. Extensive experiments on a variety of real-world bipartite networks show that BiSPM achieves a more competitive and more robust performance in comparison with the state-of-art link prediction methods in bipartite networks.