Monte Carlo Based Incremental PageRank on Evolving Graphs

Abstract

Computing PageRank for enormous and frequently evolving real-world network consumes sizable resource and comes with large computational overhead. To address this problem, IMCPR, an incremental PageRank algorithm based on Monte Carlo method is proposed in this paper. IMCPR computes PageRank scores via updating previous results accumulatively according to the changed part of network, instead of recomputing from scratch. IMCPR effectively improves the performance and brings no additional storage overhead. Theoretical analysis shows that the time complexity of IMCPR to update PageRank scores for a network with m changed nodes and n changed edges is O((m+n/c)/c), where c is reset probability. It takes O(1) works to update PageRank scores as inserting/removing a node or edge. The time complexity of IMCPR is better than other existing state-of-art algorithms for most real-world graphs. We evaluate IMCPR with real-world networks from different backgrounds upon Hama, a distributed platform. Experiments demonstrate that IMCPR obtains PageRank scores with equal (or even higher) accuracy as the baseline Monte Carlo based PageRank algorithm and reduces the amount of computation significantly compared to other existing incremental algorithm.