Combining metadata and co-citations for recommending related papers

Abstract

Identification of relevant documents is performed to keep track of the state-of-the-art methods and relieson research paper recommender systems. The proposed approaches for these systems can be classified into categorieslike content-based, collaborative filtering-based, and bibliographic information-based approaches. The content-basedapproaches exploit the full text of articles and provide more promising results than other approaches. However, mostcontent is not freely available because of subscription requirements. Therefore, the scope of content-based approachesis limited. In such scenarios, the best possible alternative could be the exploitation of other openly available resources.Therefore, this research explores the possible use of metadata and bibliographic information to find related articles.The approach incorporates metadata with co-citations to find and rank related articles against a query paper. Thesimilarity score of metadata fields is calculated and combined with co-citations. The proposed approach is evaluated ona newly constructed dataset of 5116 articles. The benchmark ranking against each co-cited document set is established byapplying Jensen–Shannon divergence (JSD) and results are evaluated with the state-of-the-art content-based approachin terms of normalized discounted cumulative gain (NDCG). The state-of-the-art content-based approach achieved anNDCG score of 0.86 while the traditional co-citation-based approach scored 0.72. The presented method achieved NDCGscores of 0.73, 0.77, and 0.78 by incorporating the title, co-citation and title, and abstract, respectively, whereas thehighest NDCG score of 0.77 was achieved by combining co-citations with metadata. However, better results are achievedby incorporating the title and abstract with NDCG score of 0.81. Therefore, it can be concluded that the proposedapproach could be a better alternative in cases where content is unavailable.