How can we succinctly describe a large, dynamic graph over time? Given a large dynamic graph, can we �find "important" patterns that evolve over time, so that we can easily summarize and visualize the graph? In real life, these patterns signify interaction between nodes over time -- for example, how the network traffic of a bank changes during the day, how calling patterns change season over season, or how people watch different genre of movies over different times of the year. Our work focuses on the problem of how to find and rank these patterns. To this end, we formalize this problem as minimizing the encoding cost in a data compression paradigm and propose Mango, an effective heuristic for discovering evolving patterns in dynamic graphs. We then apply our method to both synthetic and real datasets to show that Mango is able to summarize dynamic graphs by meaningful static and temporal patterns.
ECMLPKDD Workshop on Graph Embedding and Mining (GEM)