On the feasibility of low-rank approximation for personalized PageRank

Personalized PageRank expresses backlink-based page quality around user-selected pages in a similar way to PageRank over the entire Web. Algorithms for computing personalized PageRank on the fly are either limited to a restricted choice of page selection or believed to behave well only on sparser regions of the Web. In this paper we show the feasibility of computing personalized PageRank by a k < 1000 low-rank approximation of the Page-Rank transition matrix; by our algorithm we may compute an approximate personalized Page-Rank by multiplying an n x k, a k x n matrix and the n-dimensional personalization vector. Since low-rank approximations are accurate on dense regions, we hope that our technique will combine well with known algorithms.