Fast PageRank Computation Based on Network Decomposition and DAG Structure

Abstract

PageRank has been widely used for the problem of evaluating the importance of data in many applications, such as Web science, information systems, and social network analysis. As vast amounts of data generate, the development of efficient PageRank computation is a vibrant area of contemporary research. In this paper, we propose a fast PageRank computation method based on network decomposition and the structure of directed acyclic graph (DAG). A network decomposition technique is first introduced to decompose the original network into three parts, including a general sub-network and two sub-DAGs. Based on the acyclic characteristic, we demonstrate that these two sub-DAGs can be theoretically lumped into two single nodes by a similarity transformation. Then, the PageRank problem on the original network is transformed into a PageRank problem on a much smaller network, and the full PageRank vector can be easily recovered from the result of new problem. As the time taken by the estimation of PageRank is directly proportional to the network size, the proposed method achieves an improved time complexity and is more efficient as the size of two sub-DAGs increases. Experimental results on real data sets show that our method provides a significant speedup compared with existing alternatives.

Keywords

• PageRank
• link analysis
• DAG structure
• network decomposition