A machine learning framework for early warning prediction of student success using privacy preserving synthetic educational data

The early prediction of student success through machine learning holds transformative potential for improving educational outcomes. However, the development of robust predictive models is fundamentally constrained by the scarcity of large-scale, high-quality educational datasets and privacy regulations. This paper introduces a machine learning framework that addresses these challenges through privacy-preserving synthetic educational data, with the key insight that synthetic data provides maximum value as a pre-training foundation rather than a complete replacement for real data. The framework is built upon SynEdu-HEDL, a novel synthetic dataset comprising 20,000 students across 180 courses with six interconnected tables. The dataset was generated using a hybrid rule-based and probabilistic approach that provides strong privacy guarantees (no real student records are used). Using this synthetic foundation, we evaluate predictive architectures including LSTM networks with attention, transformers, graph neural networks, and a novel hybrid LSTM-GNN architecture. Experimental results on the real-world OULAD dataset demonstrate that while direct transfer from synthetic to real data yields limited performance (AUC-ROC = 0.714), fine-tuning with only 5% of real data achieves an AUC-ROC of 0.781-a 12.7% relative improvement over training from scratch on the same limited real data (0.693). With 20% real data, fine-tuning reaches AUC-ROC of 0.831, approaching the full-data upper bound of 0.842. This synthetic pre-training benefit represents the central practical contribution of this work. Temporal analysis reveals that meaningful early warnings can be generated by week four, with the LSTM-GNN model correctly identifying 47.3% of students who would ultimately drop out. Fairness evaluation identifies acceptable disparities for gender and program type but meaningful differences for first-generation students (disparate impact ratio 0.84 on OULAD), which is successfully mitigated with only 1.8% reduction in predictive performance. Domain gap analysis reveals that performance degradation stems primarily from differences in forum participation (KS=0.34) and login frequency (KS=0.21) between synthetic and real data. The SynEdu-HEDL dataset and all code are publicly released. The framework provides validated evidence that synthetic educational data offers the greatest practical value when used for pre-training, enabling institutions with limited historical data to develop effective early warning systems without compromising student privacy.