ARISE-GNN: Graph Neural Representation Learning for Discovering Potential Rising-Star Candidates in Scholarly Collaboration Networks

Abstract

Identifying authors exhibiting characteristics associated with emerging scholarly influence at an early career stage is a critical challenge in scientometrics and research evaluation. Existing approaches predominantly rely on isolated indicators such as citation count, H-index, and network-based metrics (e.g., PageRank), which inadequately capture the complex and multi-relational nature of scientific influence. To address this limitation, this study proposes ARISE-GNN (Author Representation Learning for Identifying Scholarly Emergence), a graph-based unsupervised framework for discovering authors exhibiting emerging influence-related characteristics within scholarly collaboration networks. A collaboration network is constructed from co-authorship data originally represented in tabular form, enabling the transformation of bibliometric information into a relational graph structure. The framework integrates diverse bibliometric features with network topology to facilitate information propagation across authors. Multiple Graph Neural Network (GNN) architectures, including GraphSAGE, Graph Autoencoders (GAE), and Deep Graph Infomax (DGI), are employed for topology-aware representation learning. In particular, a hybrid GraphSAGE+DGI model is introduced to jointly capture local structural dependencies and global graph semantics through contrastive self-supervised learning while effectively addressing label scarcity. The key novelty lies in the integration of heterogeneous bibliometric features with self-supervised graph representation learning for label-free discovery of emerging scholarly influence patterns. Experimental results demonstrate superior representation quality, achieving a Silhouette Score of 0.67, a Davies–Bouldin Index of 0.94, and an embedding standard deviation of 0.71, indicating improved cluster compactness, separation, and embedding dispersion. The proposed framework provides a scalable and effective unsupervised approach for identifying authors with characteristics associated with emerging scholarly influence based on graph representations and bibliometric indicators. The resulting rankings should be interpreted as identifying potential rising-star candidates rather than definitively confirmed future high-impact researchers, as predictive temporal validation is beyond the scope of the present study.