Personalized PageRank Solution Paths.
Year: 2015
Type: preprint
Abstract: Personalized PageRank vectors used for many community detection and graph diffusion problems have a subtle dependence on a parameter epsilon that controls their accuracy. This parameter governs the sparsity of the solution and can be interpreted as a regularization parameter. We study algorithms to estimate the solution path as a function of the sparsity and propose two methods for this task. The first computes a full solution path and we prove it remains localized in the graph for fast runtimes. Using this method, we propose a PageRank solution path plot to diagnose new aspects of the behavior of personalized PageRank. The second method is a faster approximation to the solution path on a grid of logarithmically-spaced values that uses an interesting application of bucket sort to make the process efficient. We demonstrate that both of these algorithms are fast and local on large networks.
Source: arXiv (Cornell University)
Authors: Kyle Kloster, David F. Gleich
Language: English
Subfield: Statistical and Nonlinear Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W1514538974