A Deep Learning Framework for Dynamic Network Link Prediction

Link prediction, which involves predicting potential relations between nodes in networks, has long been a challenge in network science. Most studies have focused on link prediction of static networks, while real-world networks constantly evolve with the addition and removal of nodes and links. More and more attention has been drawn to dynamic network link prediction since it can better capture the evolution nature of networks, but most algorithms still fail to achieve satisfactory prediction accuracy. A novel deep learning model to predict dynamic links end to end is proposed in this paper, motivated by the excellent performance of Long Short-Term Memory (LSTM) in processing time series. It has the ability to handle long-term prediction problems and is suitable for networks of different scales with fine-tuned structure. It is believed to be the first time that LSTM, along with an encoder–decoder architecture, is applied to link prediction in dynamic networks. The new model is capable of automatically learning structural and temporal features in a unified framework, which allows it to predict links that have never appeared in the network before. Evident from the extensive experiments, our method surpasses the newly proposed dynamic network link prediction methods and achieves state-of-the-art results.