Improved Link-Based Algorithms for Ranking Web Pages

Abstract

Several link-based algorithms, such as PageRank [7], HITS [4] and SALSA [5], have been developed to evaluate the popularity of web pages. These algorithms can be interpreted as computing the steady-state distribution of various Markov processes over web pages. The PageRank and HITS algorithms tend to over-rank tightly interlinked collections of pages, such as well-organized message boards. We show that this effect can be alleviated using a number of modifications to the underlying Markov process. Specifically, rather than weight all outlinks from a given page equally, greater weight is given to links between pages that are, in other respects, further off in the web, and less weight is given to links between pages that are nearby. We have experimented with a number of variants of this idea, using a number of different measures of ”distance” in the Web, and a number of different weighting schemes. We show that these revised algorithms often do avoid the over-ranking problem and give better overall rankings.

This work is supported by NSF grant #IIS-0097537. There is an extended version of this paper as NYU computer science dept. technical report TR2003-846, online at http://csdocs.cs.nyu.edu/Dienst/UI/2.0/Describe/ncstrl.nyu cs/TR2003-846.