Experimental Evaluation of Dynamic Graph Clustering Algorithms

Graph clustering is concerned with identifying the group structure of networks. Existing static methods, however, are not the best choice for evolving networks with evolving group structures. We discuss dynamic versions of existing clustering algorithms which maintain and modify a clustering over time rather than recompute it from scratch. We developed an extensible software framework for the evaluation of these algorithms, and present experimental results on real-world and synthetic graph instances. Our focus on clustering quality, clustering smoothness, and runtime. We conclude that dynamically maintaining a clustering on an evolving graph is superior in terms of all criteria. We demonstrate that dynamic algorithms are able to react quickly and appropriately to changes in the cluster structure. Our results allow us to give sound recommendations for the choice of an algorithm.