A Sequential and Scalable Approach to Community Detection in Dynamic Graphs

Andre Beckus, George K. Atia

Abstract: We study a sequential sketch-based approach for the clustering of time-evolving graphs. We present a dynamic extension to the static Stochastic Block Model, which accommodates growing and shrinking graphs, as well as the movement of nodes between clusters. We then propose an online algorithm which constructs and maintains a small sketch consisting of nodes sampled from the full graph. The sketch is clustered and a retrieval algorithm is used to infer cluster membership of nodes in each successive graph snapshot. We demonstrate that the use of a small sketch not only improves computational complexity, but also improves the success rate when sketches are properly proportioned. We present a sampling method which chooses nodes according to node degree, whereby very small clusters can be successfully tracked.