Perron vector optimization applied to search engines

In the last years, Googleʼs PageRank optimization problems have been extensively studied. In that case, the ranking is given by the invariant measure of a stochastic matrix. In this paper, we consider the more general situation in which the ranking is determined by the Perron eigenvector of a nonneg...

Full description

Saved in:
Bibliographic Details
Published in:Applied numerical mathematics Vol. 75; pp. 77 - 99
Main Author: Fercoq, Olivier
Format: Journal Article
Language:English
Published: Elsevier B.V 01.01.2014
Subjects:
Algorithms
Invariants
Mathematical analysis
Optimization
Perron–Frobenius theory
Ranking
Scalable algorithms
Search engines
Stochasticity
Strategy
Search engines
Scalable algorithms
Optimization
Perron–Frobenius theory
ISSN:0168-9274, 1873-5460
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In the last years, Googleʼs PageRank optimization problems have been extensively studied. In that case, the ranking is given by the invariant measure of a stochastic matrix. In this paper, we consider the more general situation in which the ranking is determined by the Perron eigenvector of a nonnegative, but not necessarily stochastic, matrix, in order to cover Kleinbergʼs HITS algorithm. We also give some results for Tomlinʼs HOTS algorithm. The problem consists then in finding an optimal outlink strategy subject to design constraints and for a given search engine. We study the relaxed versions of these problems, which means that we should accept weighted hyperlinks. We provide an efficient algorithm for the computation of the matrix of partial derivatives of the criterion, that uses the low rank property of this matrix. We give a scalable algorithm that couples gradient and power iterations and gives a local minimum of the Perron vector optimization problem. We prove convergence by considering it as an approximate gradient method. We then show that optimal linkage strategies of HITS and HOTS optimization problems satisfy a threshold property. We report numerical results on fragments of the real web graph for these search engine optimization problems.
Bibliography:ObjectType-Article-2
SourceType-Scholarly Journals-1
ObjectType-Feature-1
content type line 23
ISSN:0168-9274
1873-5460
DOI:10.1016/j.apnum.2012.12.006