Recommending Scientific Papers: A Survey and A Hybrid Approach

Abstract

Paper recommendation systems are important tools for helping scholars and researchers discover relevant and interesting papers in the face of the growing volume of published research. In this study, we conduct a survey of paper recommendation approaches and evaluate the performance of our own implementation, the ExtendedPaperVeTo recommender. Our survey covers a range of approaches and studies, and identifies common trends and challenges in the field. Our implementation evaluation compares the performance of the ExtendedPaperVeTo approach to two other approaches, the MongoFTS and PaperVeTo approaches, using various evaluation metrics. Our findings indicate that hybrid approaches are the most common type of paper recommendation approach (PRA) used in the reviewed studies, followed by content-based filtering, graph-based approaches, and collaborative filtering. However, we also identify several challenges and limitations in the current state of the field, including the lack of reproducibility and scalability of many approaches, the limited consideration of the operator’s perspective and user characteristics, and the reliance on offline evaluations. In our implementation evaluation, we observe that the ExtendedPaperVeTo approach was slightly outperformed by the MongoFTS approach in terms of NDCG and AR scores, but performed better than the PaperVeTo approach. In the conclusion, we discuss potential directions for future research that could address the challenges and limitations identified in our survey and implementation evaluation.