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

Area of Science:

• Educational Technology
• Data Science
• Artificial Intelligence

Background:

• Traditional education evaluation methods struggle with data silos, leading to local overfitting and failed dynamic correlation modeling.
• Cross-school educational data presents challenges due to heterogeneous distributions and privacy concerns.

Purpose of the Study:

• To introduce a novel evaluation framework combining space-time graph diffusion model (STG-DM) and federated contrastive learning (FedCL) for collaborative optimization in cross-school education.
• To develop a system that accurately models the space-time evolution of multimodal educational behaviors while overcoming data silo limitations.

Main Methods:

• Developed a thermodynamically driven space-time diffusion equation and an adaptive graph convolution mechanism.
• Integrated STG-DM with FedCL to enable collaborative optimization and address privacy leakage risks.
• Implemented a multimodal real-time analysis engine for space-time heatmap rendering and collaborative decision-making.

Main Results:

• Reduced average absolute error (MAE) by 18.7% in cross-regional education strategy prediction compared to conventional space-time models.
• Successfully reduced privacy leakage risk (ε) to below 1.5 in a heterogeneous data distribution across 30 universities.
• Achieved balanced optimization of cross-school model generalization performance with a system response delay under two seconds for 100,000-level nodes.

Conclusions:

• The novel framework effectively overcomes limitations of traditional methods in cross-school education evaluation.
• The system provides efficient and reliable data intelligence tools for education managers, enhancing collaborative decision-making.
• The integration of STG-DM and FedCL offers a robust solution for privacy-preserving, accurate, and scalable educational data analysis.