Evaluation of kernel-based link analysis measures on research paper recommendation

Abstract

We compare various kernel-based link analysis measures on graph nodes to evaluate their utility as a research paper recommendation system. The compared measures include the Kandola et al.'s von Neumann kernel, its extension that takes communities into account, and Smola and Kondor's regularized Laplacian. Chebotarev and Shamis' matrix forest-based algorithm, Kleinberg's HITS authority ranking, and classic co-citation coupling are also evaluated. The experimental result shows that kernel-based methods outperform HITS and co-citation coupling, with the community-based von Neumann kernel achieving the highest score.