Research on cooperative relationship prediction based on transformer dynamic graph modeling and comparative learning

Abstract

The goal of this study is to develop an accurate and efficient framework for predicting scientific research cooperation relationships in dynamic academic networks. Existing temporal graph learning models often struggle with sparse and heterogeneous collaboration data, limiting their applicability in real-world academic settings. To address this challenge, we propose the transformer-based contrastive learning (TCL) model, which integrates a topology-aware transformer with a dual-stream encoder to capture structural and temporal dependencies in evolving collaboration graphs jointly. We construct dynamic network snapshots from two large-scale datasets: National Key R&D Project (NKPD) collaboration networks and citation network data (CND). Comparative experiments against state-of-the-art baselines, including DySAT, TGAT, TGN, and CLDG, demonstrate that TCL achieves the highest accuracy-up to 95.3%area under the ROC curve (AUC)—while reducing training time by 21%. Furthermore, our analysis reveals that network sparsity and time-slice granularity significantly influence prediction performance, with annual slices outperforming half-year slices by 35.8% in AUC. The impact of this work lies in providing a robust and scalable tool for forecasting research partnerships, enabling funding agencies, institutions, and policymakers to identify high-potential collaborations earlier. This contributes to accelerating interdisciplinary innovation and enhancing the strategic allocation of research resources.