Recommending collaboration links in coauthorship networksŽontar, Luka (Avtor) Curk, Tomaž (Mentor) graph neural networkslink predictionhomogeneous networksbibliography miningscientometricsIn the evolving landscape of academic research, interdisciplinary and cross-institutional collaboration has become critical. However, identifying valuable partnerships remains a complex task. This thesis introduces a recommender system for suggesting new academic collaborations within coauthorship networks, focusing on promoting cross-institutional research within the EUTOPIA alliance. We construct a large-scale coauthorship dataset from bibliographic sources such as Elsevier, Crossref, and ORCID, and model it using graph neural networks. Our work involves an ablation study of various GNN backbones, including LightGCN, GraphSAGE, and attention-based models, combined with different loss functions and node feature sets. The final model uses weighted temporal embeddings of article content and achieves MRR@10=19.0% ± 0.3% and HitRate@10=36.4% ± 0.5%, outperforming the baseline methods. Although attempts to incorporate keyword popularity as a feature were inconclusive, our work highlights the potential of GNN-based recommender systems to enhance research collaboration networks. In addition, we provide a detailed analysis of collaboration dynamics, highlighting the importance of new research collaborations and investigating how research trends, research interests, and lead authors affect the occurrence of new collaborations.20252025-11-18 10:20:01Magistrsko delo/naloga176013sl