Efficient PageRank Computation via Distributed Algorithms with Web Clustering

TL;DR

This paper introduces a novel distributed PageRank algorithm that uses web clustering and gossip-based randomization to achieve exponential convergence with reduced communication, demonstrated on real web data.

Contribution

It presents a new reinterpretation of PageRank leading to algorithms with exponential convergence, incorporating clustering to improve efficiency and reduce communication.

Findings

Exponential convergence rates achieved

Clustering-based scheme accelerates convergence

Reduced communication compared to existing methods

Abstract

PageRank is a well-known centrality measure for the web used in search engines, representing the importance of each web page. In this paper, we follow the line of recent research on the development of distributed algorithms for computation of PageRank, where each page computes its own PageRank value by interacting with pages connected over hyperlinks. Our approach is novel in that it is based on a reinterpretation of PageRank, which leads us to a set of algorithms with exponential convergence rates. We first employ gossip-type randomization for the page selections in the update iterations. Then, the algorithms are generalized to deterministic ones, allowing simultaneous updates by multiple pages. Finally, based on these algorithms, we propose a clustering-based scheme, in which groups of pages make updates by locally interacting among themselves many times to expedite the convergence. In comparison with other existing techniques, significant advantages can be exhibited in their convergence performance, as demonstrated via numerical examples using real web data, and also in the limited amount of communication required among pages.