Recommending research articles using citation data
Published date | 16 November 2015 |
Pages | 597-609 |
DOI | https://doi.org/10.1108/LHT-06-2015-0063 |
Date | 16 November 2015 |
Author | Andre Vellino |
Subject Matter | Library & information science,Librarianship/library management,Library technology |
Recommending research
articles using citation data
Andre Vellino
University of Ottawa, Ottawa, Canada
Abstract
Purpose –The purpose of this paper is to present an empirical comparison between the
recommendations generated by a citation-based recommender for research articles in a digital library
with those produced by a user-based recommender (ExLibris “bX”).
Design/methodology/approach –For these computer experiments 9,453 articles were randomly
selected from among 6.6 M articles in a digital library as starting points for generating
recommendations. The same seed articles were used to generate recommendations in both
recommender systems and the resulting recommendations were compared according to the “semantic
distance”between the seed articles and the recommended ones, the coverage of the recommendations
and the spread in publication dates between the seed and the resulting recommendations.
Findings –Out of the 9,453 test runs, the recommendation coverage was 30 per cent for the user-based
recommender vs 24 per cent for the citation-based one. Only 12 per cent of seed articles produced
recommendations with both recommenders and none of the recommended articles were the same. Both
recommenders yielded recommendations with about the same semantic distance between the seed
article and the recommended articles. The average differences between the publication dates of the
recommended articles and the seed articles is dramatically greater for the citation-based recommender
(+7.6 years) compared with the forward-looking user-based recommender.
Originality/value –This paper reports on the only known empirical comparison between the Ex
Librix “bX”recommendation system and a citation-based collaborative recommendation system.
It extends prior preliminary findings with a larger data set and with an analysis of the publication
dates of recommendations for each system.
Keywords Digital libraries, Library services, Computer applications, Citation analysis,
Collaborative filtering, Recommender systems
Paper type Research paper
1. Introduction
Despite the relatively slow acceptance of recommender systems technology in library
settings (Wakeling et al., 2012), libraries and digital libraries especially, continue to be
an important application domain for recommender systems. The rapid growth of open
access scholarly research and an increasingly long tail of rarely cited articles, puts an
onus on intelligent discovery tools to suggest literature that is not necessarily keyword
related to the items that the user has already found with search terms and yet remains
topically relevant while offering some degree of serendipity.
The decision to either develop an in-house recommender system or to purchase a
recommender system provided by a commercial third party is complex and depends
not only on the choice of underlying recommender technology but also on its
effectiveness for the user community, the quality of the user interface and the
trustworthiness of the recommendations. This was the situation facing the Canada
Institute for Scientific and Technical Information (CISTI) (now the National Research
Council’s National Science Library) in 2012 and the initial motivation for this study: to
empirically compare the recommendations produced by two recommender systems
employing different data sources and algorithms.
Library Hi Tech
Vol. 33 No. 4, 2015
pp. 597-609
©Emerald Group Publis hing Limited
0737-8831
DOI 10.1108/LHT-06-2015-0063
Received 7 June 2015
Revised 5 September 2015
Accepted 10 September 2015
The current issue and full text archive of this journal is available on Emerald Insight at:
www.emeraldinsight.com/0737-8831.htm
597
Recommending
research articles
To continue reading
Request your trial