• recommendation system
• research artefact
• graph learning
• digital library
• deep learning
• comparative study

The task of recommendation in scientific domain is to provide researchers with a list of research entities that might be in their interests. Recommendation systems for science have become very popular due to information overload and gained a lot of research attention. Though the major focus is on developing research paper recommendation systems, other research artefacts are left untouched while they are worth knowing artefacts. This work introduces a framework through which researchers are provided with research artefacts including publications, datasets, and softwares relevant to their profiles. To this end, we exploit the OpenAIRE research graph data for the knowledge source of the framework. Then, we carefully analyse possible implementations of this framework with different recommendation approaches. In particular, a content-based model (i.e., SPECTER), a graph-based model (i.e., Node2Vec) and two hybrid models (i.e., Graph Convolutional Network and GraphSAGE) combining the content and graph structural information were implemented and compared. These recommendation models are built upon prior work on deep learning. The models yield embeddings of items and top-Q recommendation list is generated based on the similarity calculated with these embeddings. Finally, a comparative study has been conducted on the experimental results obtained for the different models. Results demonstrate the superiority of a hybrid approach using GraphSAGE model over others as it achieved higher hit rate and recall for link prediction tasks and faster training time.