Graph Embedding Framework for Link Prediction and Vertex Behavior Modeling in Temporal Social Networks

We present a novel framework in which the link prediction problem in temporal social networks is formulated as trajectory prediction in a continuous space. Four major modules constitute this framework: (1) graph embedding: the discrete space of graphs is mapped into a continuous space while preserving distances for a given graph kernel; (2) manifold alignment: graph embeddings corresponding to different time points are aligned to achieve low variance trajectories for a more reliable prediction; (3) trajectory prediction: the temporal graphs form a time series, to which various prediction models (e.g. regression) can be applied; (4) graph reconstruction from the predicted graph embedding which is invariant against scale, translation and rotation. Furthermore, this framework enables an innovative way to analyze temporal graph vertex behaviors and visualization. Extensive preliminary results on real world data sets demonstrate the promises of this proposed approach.