Representation Learning of Dynamic Networks

Dynamic networks encode the time-dependent relationships among individuals in a population. The changes in the population are not only happening at observed and discrete points but also continuously over time. Most existing learning methods can only handle a sequence of discrete snapshots. This work introduces a novel representation learning method to improve both interpretability and predictive power for network dynamics. The approach of using functional data analysis naturally defines the learning components over continuous supports while enabling attribute learning, community detection, and link prediction. The method is able to pool information across different time snapshots of the network and enable evaluation/interpolation of edge probabilities between observation times via the fitted functional representations. In addition, we provide a scalable algorithm to estimate the time-dependent representations of nodes in the network. Both simulation studies and real-world applications demonstrate the effectiveness of the proposed method. For example, we apply our method to dynamic social networks in ant colonies, uncovering meaningful patterns in interactions and role transitions. The node representations uncover biologically plausible role transitions and achieve state-of-the-art link recovery accuracy. This work provides a statistical framework balancing representation learning capacity with interpretability, offering insights into dynamic network structures.