Big Data Prediction of Scientific Collaboration Networks Based on Spatiotemporal Dynamics

The spatiotemporal dynamic evolution of scientific collaboration networks is crucial for understanding academic innovation patterns and optimizing research resource allocation. However, existing models have limitations in capturing these dynamics accurately. To address this, we propose a novel Spatio-Temporal Graph Convolutional Network (ST-GCN) framework. Firstly, we construct a multidimensional dynamic network model incorporating author co-authorship networks , keyword co-occurrence networks, and institutional collaboration networks. Edges within these networks are dynamically weighted based on the composite journal impact factors, updated with a temporal decay function. Secondly, a spatio-temporal graph convolution module based on neighborhood partitioning is designed, dividing node neighborhoods into self-connections, frequent collaborations, and infrequent collaborations, effectively capturing cross-period collaboration evolution. Additionally, a predictive deep learning model integrating multiple features, including journal impact factors, geographic encoding of institutions, and historical collaboration strength, employs a hierarchical attention mechanism to optimize feature weighting. Experimental results using data collected from CNKI involving 1786 publications by top-performing computer science scholars demonstrate that the proposed ST-GCN model effectively captures the spatiotemporal dynamics of scientific collaboration networks. This research not only enriches the theoretical understanding of collaborative networks but also provides robust practical guidance for optimizing research resource allocation and predicting future collaborative relationships.