Continuous Similarity Computation over Streaming Graphs

Large network analysis is a very important topic in data mining. A significant body of work in the area studies the problem of node similarity. One way to express node similarity is to associate with each node the set of 1-hop neighbors and compute the Jaccard similarity between these sets. This information can be used subsequently for more complex operations like link prediction, clustering or dense subgraph discovery. In this work, we study algorithms to monitor the result of a similarity join between nodes continuously, assuming a sliding window accommodating graph edges. Since the arrival of a new edge or the expiration of an existing one may change the similarity between several node pairs, the challenge is to maintain the similarity join result as efficiently as possible. Our theoretical study is validated by a thorough experimental evaluation, based on real-world as well as synthetically generated graphs, demonstrating the superiority of the proposed technique in comparison to baseline approaches.