DNGR: Deep Neural Graph-Based Recommendation System for Scholarly Paper Retrieval

M Sathya, Arulmurugan Ramu, K Velkumar, M. Bhavani

2025

Full text Open access: Gold Metadata (CSL-JSON)
DOI 10.14419/dzzstd42
Data confidence 70 medium
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

With the exponential growth of scholarly publications, identifying relevant literature has ‎become increasingly challenging for researchers. Traditional recommendation methods—such ‎as collaborative filtering, content-based filtering, and citation analysis—often struggle to scale ‎or capture deep semantic and structural relationships. In this study, we introduce DNGR ‎‎(Deep Neural Graph-based Recommendation), a novel model that combines graph neural ‎networks (GNNs) with SciBERT-based semantic embeddings to enhance scholarly paper ‎recommendations. DNGR constructs a heterogeneous academic knowledge graph ‎incorporating citation links, author collaborations, and topical associations. Each paper is ‎represented by both its contextual semantics and structural position in the academic network. ‎We evaluate DNGR on the AMiner v12 DBLP-Citation Network, comprising over 4.8 million ‎papers and 45 million citation edges. Experiments using standard information retrieval ‎metrics—Mean Average Precision (MAP), Mean Reciprocal Rank (MRR), and Normalized ‎Discounted Cumulative Gain (NDCG)—demonstrate that DNGR outperforms state-of-the-art-‎art baselines, achieving up to a 92% improvement in recommendation accuracy over methods ‎like MSCN and Google Scholar. Our results highlight the potential of integrating deep ‎contextual and graph-based signals for scalable and accurate citation recommendation. We ‎also discuss limitations, ethical considerations, and potential extensions to multi-disciplinary ‎datasets‎.