A spatio-temporal graph diffusion and federated contrastive learning framework for cross-institutional educational evaluation

This study introduces a novel evaluation framework that combines a space-time graph diffusion model (STG-DM) and federated contrastive learning (FedCL) to address collaborative optimization challenges in cross-school education evaluation. This integration enables the creation of a thermodynamically driven space-time diffusion equation and an adaptive graph convolution mechanism, facilitating accurate modeling of the space-time evolution of multimodal educational behaviors. It effectively overcomes the shortcomings of traditional methods, which often suffer from local overfitting and dynamic correlation modeling failures due to data silos. The graph diffusion operator, constrained by non-equilibrium thermodynamic principles, has proven to enhance the prediction accuracy of cross-regional education strategies, reducing the average absolute error (MAE) by 18.7% compared to conventional space-time models. In the context of heterogeneous data distribution across 30 universities, the system successfully reduces the privacy leakage risk (ε) to below 1.5, while simultaneously achieving balanced optimization of cross-school model generalization performance. The lightweight evaluation system developed includes a multimodal real-time analysis engine that enables space-time heatmap rendering and collaborative decision-making for 100,000-level nodes, with a system response delay of less than two seconds. This provides education managers with efficient and reliable data intelligence tools.