Parallel index-based structural graph clustering and its approximations

Abstract

SCAN (structural clustering algorithm for networks) is a well-known approach for graph clustering. Sequential versions of SCAN are prohibitively slow on large graphs, however. Existing parallel versions of SCAN, on the other hand, can cluster graphs relatively quickly on a particular setting of SCAN parameters, but do not effectively share work among queries on different parameter settings. Because users of SCAN need to test several parameter settings in order to find a good clustering, it can be worthwhile to precompute an index to speed up later queries. To that end, this thesis presents a parallelization of GS*-Index, an existing index-based SCAN algorithm. The parallelized algorithm is work-efficient and achieves logarithmic span for both constructing the index and running clustering queries. We describe an implementation of our algorithm and test it on several real-world large graphs, with the largest graph having 1.8 billion edges. On a machine with 48 cores and 2-way hyper-threading, our parallel index construction achieves 50-151x speedup over the construction of GS*-Index. In fact, our index construction algorithm is faster than GS*-Index even when running our algorithm sequentially. Our parallel index query implementation achieves 5-32x speedup over queries on GS*- Index across a range of SCAN parameter values, and our implementation is also faster than ppSCAN, the fastest existing parallel SCAN algorithm, on all tested parameter values. We also explore how locality-sensitive hashing can speed up index construction by approximating the similarity scores between vertices, the computation of which is the most time-consuming aspect of SCAN. Our experiments show that this technique can achieve meaningful speedups on denser graphs without large sacrifices in clustering quality.