Benchmarking GATv2, EvolveGCN, and Node2Vec for Next-Year Co-Authorship Prediction on arXiv

This study benchmarks three representative paradigms, Node2Vec, GATv2, and EvolveGCN, for next-year co-authorship prediction on a large-scale dynamic collaboration network constructed from arXiv cs.CL metadata from January 1, 2010 to December 30, 2025. To mitigate topology bias caused by homonymous authors, community-based author disambiguation is performed using the Leiden algorithm and an Ego-Splitting strategy. SciBERT-based semantic representations are incorporated with an exponential time-decay mechanism to model the temporal drift of research interests. Experimental results show that GATv2 achieves the best predictive performance with a Test AUC of 0.9089, whereas EvolveGCN provides a lightweight and efficient alternative with 0.66 million parameters and substantially shorter training time than GATv2. In addition, Preferential Attachment remains competitive on sparse graphs, while semantic and spatial fusion models better capture complex collaboration patterns. Overall, the results quantify the accuracy and efficiency trade-off and provide practical guidance for model selection in large-scale dynamic academic link prediction.