Expert Finding Revisited: A Uniform Exploration of Methods

Le résumé fourni par la source

The goal of expert search is to identify individuals with knowledge or skills in a specific field or for a specific problem. In the context of academia, expert search is enabled by resources such as publications, profiles, and authorship networks; with numerous methods proposed. However, they have been tested on diverse and inconsistent collections, making comparison difficult. Moreover, some of these collections reflect a narrow interpretation of the problem. Our focus in this paper is a fresh examination of expert search, which we pursue by re-evaluation of four previous methods under consistent experimental conditions. We explore those models by factorially varying their similarity functions, and by factorially varying their uses of evidence. Our experiments make use of collections from previous experiments and new test sets generated by a methodology we propose, the latter created in a straightforward and replicable way from publicly available data, with the benefit of not needing relevance judgments to be formed. Key findings from our results are that no one previous method is strongest overall and that relative performance depends on the search need: one family of methods is better for area-based search, while another is superior when searching for experts working on specific topics. We thus conclude that many previous claims about individual expert search methods are not broadly supported, and hence that the problem is richer than previous individual proposals have suggested.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.