Link Prediction in Time Varying Social Networks

Predicting new links in complex networks can have a large societal impact. In fact, many systems be modeled through networks, and the meaning of depend on system itself. For instance, social where nodes are users, represent relationships (such as acquaintance, friendship, etc.), whereas information spreading users content interactions, diffusion, etc. However, while approaches involve machine learning-based algorithms, just most recent ones account for topology network, e.g., geometric deep learning techniques to learn graphs, them do not temporal dynamics network but train snapshots at given time. this paper, we aim explore Temporal Graph Networks (TGN), Representation Learning-based approach that natively supports dynamic graphs assigns each event (link) timestamp. particular, investigate how TGN behaves when trained under different granularity or with various aggregation inductive transductive link prediction problem real such Twitter, Wikipedia, Yelp, Reddit. We find initial setup affects data, impact depends specific network. note batch size has strong it does matter