Predicting neighbor label distributions in dynamic heterogeneous information networks

摘要

In Dynamic Heterogeneous Information Networks (DHINs), predicting neighbor label distribution is important for a variety of applications. For example, when a user changes job, the composition of the user's friends can change, hence the profession distribution of his/her social circle may change. If we can accurately predict the change of the distribution, we will be able to improve the quality of personal services for him/her. The challenges of predicting neighbor label distribution mainly come from four aspects: infinite state space of neighbor label distributions, link sparsity, the complexity of link formation preferences, and the stream of DHIN snapshots. To address these challenges, we propose a Latent Space Evolution Model (LSEM) for the prediction of neighbor label distribution, which builds a Neighbor Label Distribution Matrix (NLDM) for each type of labels of neighbors of given nodes. LSEM can predict the next NLDM by reconstructing it from two latent feature matrices estimated by their respective autoregressive models. The experiments conducted on real datasets verify the effectiveness of LSEM and the efficiency of the proposed algorithm.