Supplementary material from "Dynamic graph neural networks for open-source scientific collaboration networks"

Abstract

Scientific collaboration is a temporal process: researchers form, strengthen and let lapse collaborations over time, and external shocks perturb the network. Here, we present a reproducible benchmark of dynamic graph neural networks for link forecasting on a large open co-authorship corpus. We use OpenAlex (data snapshot of 22–23 September 2026) across six disciplines—condensed matter physics, genetics, epidemiology, organic chemistry, artificial intelligence and language and linguistics—spanning 2000–2025. Using strict temporal splits with no future-information leakage, we compare three dynamic architectures (T-GCN, DGCN and EvolveGCN-O), a temporal-attention baseline (DySAT) and three static baselines, under a rolling-origin evaluation with three temporal cuts and three random seeds. T-GCN achieves the best AUC-ROC and AUC-PR (0.777 and 0.978, respectively), outperforming the strongest static baseline by 8.7% and the next-best model excluding the scale-limited DySAT reference—a GRU-cell EvolveGCN ablation—by 4.7 AUC points. Clustering the best model's embeddings reveals three collaboration archetypes—core contributors, community bridges and peripheral specialists—that are stable across random seeds (multi-seed ARI 0.998), independent model retrainings (cross-embedding ARI 0.85), and partially across clustering methods (k-means versus agglomerative ARI 0.52; spectral clustering does not recover the partition), and exhibit substantial between-archetype mobility; they are therefore presented as exploratory. A COVID-19 case study using synthetic control and difference-in-differences illustrates how the framework can estimate associations between external shocks and cross-disciplinary collaboration; these estimates are associational, not causal, and the headline result is a placebo-validated null. All data and code are openly available.