A privacy preserving synthetic learner dataset for learning analytics in technology enhanced higher education

Area of Science:

• Educational Data Science
• Artificial Intelligence in Education
• Privacy-Preserving Machine Learning

Background:

• Technology-enhanced learning generates vast student data, presenting opportunities for learning analytics.
• Existing data sharing is limited by privacy regulations and ethical concerns, hindering research.
• A novel synthetic dataset is needed to balance data utility with privacy protection.

Purpose of the Study:

• To introduce SynEdu-HEDL, a synthetic dataset for privacy-preserving learning analytics in higher education.
• To address the tension between data utility and privacy in educational data sharing.
• To facilitate collaborative research and methodological advancements in the field.

Main Methods:

• Developed SynEdu-HEDL using a five-phase framework with conditional tabular generative adversarial networks, temporal sequence generators, and differential privacy.
• The dataset includes 20,000 synthetic student records with 85 features.
• Validated using a three-dimensional framework assessing privacy, statistical fidelity, and analytical utility.

Main Results:

• SynEdu-HEDL offers strong privacy guarantees (AUC-ROC=0.512) and preserves statistical properties (avg. Wasserstein distance=0.043).
• Models trained on SynEdu-HEDL achieve performance within 1-5% of models trained on original data.
• Transfer learning experiments showed a 24.4% performance improvement using only 10% real data.

Conclusions:

• SynEdu-HEDL provides a practical solution to educational data sharing barriers.
• The dataset supports reproducible research and establishes standards for synthetic educational data validation.
• Openly available, SynEdu-HEDL fosters community engagement in privacy-preserving learning analytics.