Towards efficient navigation in digital libraries: Leveraging popularity, semantics and communities to recommend scholarly articles

Abstract

• Our study proposes a novel algorithm, SPACE-R, which aims to improve the scalability and accuracy of large real-world academic recommendation systems. • This study incorporates the “quality” of papers by the construction of a Popularity Network, a derivative of the citation network, with importance given to author and journal impact. • The significance of a neighborhood network of a dynamic radius r has been demonstrated in the overall performance of the recommender system. • The judicious incorporation of a weighted sum of recency, quality, and semantic similarity as a personalization vector in the ranking process enhances the diversity of recommendations. • Community structure has been leveraged to get more topically similar papers as well as to reduce the search space, making the recommendation process scalable. The growing popularity of digital libraries as a medium for communicating scientific discoveries has made a large variety of research articles easily accessible. However, this constitutes a putative issue of information overloading with recommendation engines providing a compelling solution to the problem. Scientific Recommender Systems alleviate this problem by suggesting potential papers of interest to a user. For any researcher seeking developments in their field, it is important that the recommended papers are of high quality, recent and related to the field of interest, which has been largely overlooked in prior approaches. This study thus proposes a graph-based hybrid recommendation technique, SPACE-R , that amalgamates quality, semantic similarity and community structure of nodes in a citation network. The creation of a popularity network, a derivative of a citation network, in combination with a two-stage candidate selection process involving community detection and neighbourhood network identification, contributes to an improvement in the accuracy and scalability of the proposed model. The incorporation of semantic similarity achieves the necessary diversity in recommendations. Experimental evaluations on four large datasets against five baselines reveal that SPACE-R achieves an improvement of up to 45.53%, 56.76%, 49.39%, 46.84% and 78.18% in recall, precision, MRR, mAP, and response time, respectively.