Usage-based vs. Citation-based Methods for Recommending Scholarly Research Articles

There are two principal data sources for collaborative filtering recommenders in scholarly digital libraries: usage data obtained from harvesting a large, distributed collection of Open URL web logs and citation data obtained from the journal articles. This study explores the characteristics of recommendations generated by implementations of these two methods: the ‘bX’ system by ExLibris and an experimental citation-based recommender, Sarkanto. Recommendations from each system were compared according to their semantic similarity to the seed article that was used to generate them. Since the full text of the articles was not available for all the recommendations in both systems, the semantic similarity between the seed article and the recommended articles was deemed to be the semantic distance between the journals in which the articles were published. The semantic distance between journals was computed from the “semantic vectors” distance between all the terms in the full-text of the available articles in that journal and this study shows that citation-based recommendations are more semantically diverse than usagebased ones. These recommenders are complementary since most of the time, when one recommender produces recommendations the other does not.