Unsupervised Link and Unlink Prediction on Dynamic Networks.

Understanding and characterizing the process deriving the creation and dissolution of social interactions is a fundamental challenge for social network analysis. In the dynamic setting, it is essential to be able to, given the collection of link states of the network in a certain period, accurately predict the link and unlink states of the network in a future time. Addressing this task is more complicated compared to its static counterpart especially for increasingly large networks due to the prohibitive expensiveness of computational complexity. Consequently, mainstreams of current researches in unsupervised settings ignore the temporal information. Additionally, only a few approaches study on unlink prediction, which is also important to understand the evolution of social networks. In this work, we address such mining tasks by unsupervised learning, and propose a model for link and unlink prediction with temporal information (LUPT). Given a sequence of snapshots of network over time, LUPT utilizes the spectral diffusion by variants of local random walks to calculate the probability vector started from each node at each snapshot. Then, it calculates the similarity score for each of the nodes by the probability vectors of all the previous snapshots. Finally, LUPT predicts the link and unlink states by ranking the similarity scores according to the link and unlink tasks, respectively. Experiments on real-world networks demonstrate that LUPT provides superior results compared to the baseline methods in both link and unlink prediction tasks.