Abstract: 202402 bcch. This paper introduces FEEDAN, a federated learning framework designed to overcome the "data island" problem in the educational sector. By enabling collaborative model training without sharing raw student data, FEEDAN preserves privacy while achieving high analytical performance comparable to centralized systems.
Introduction: The Data Island Challenge
In the era of "Education 4.0," data-driven insights—such as predicting student dropout rates or assessing academic performance—are vital for improving learning outcomes. However, educational data is highly sensitive. Legal frameworks like the Family Educational Rights and Privacy Act (FERPA) and general privacy concerns make institutions reluctant to share data.
This creates isolated data islands: small, disconnected datasets that lack the volume and variety needed to train robust machine learning models. Traditionally, the only way to obtain "big data" results was to centralize everything in one place, which creates a massive privacy risk. This research addresses a critical deadlock: how can we gain the benefits of cross-institutional big data analysis without compromising student confidentiality?
The FEEDAN Framework
The authors propose FEEDAN (FEderated Education Data ANalysis), a framework built on the principles of Federated Learning (FL). Unlike traditional machine learning, which requires data to be moved to a central server, FL keeps the data local.
The Core Mechanism
In the FEEDAN architecture, a central server distributes a "global model" to participating institutions. Each institution trains this model locally on its own private data. Once local training is complete, they send only the model updates (such as weights and loss values) back to an Aggregation Server. The central server then uses algorithms like FedAvg to combine these local updates into an improved global model, which is redistributed for the next round of training.
Implementation Paradigms
FEEDAN is designed to handle two primary educational scenarios:
- Horizontal Federated Learning: Applied when institutions have different students but record similar information (e.g., two different universities tracking standardized grades).
- Vertical Federated Learning: Applied when different institutions have the same students but different types of data (e.g., combining a university's grade records with a MOOC platform's clickstream data).
To minimize the "destructive change" required for adoption, the framework is compatible with existing IT infrastructure, utilizing standard SQL databases, firewalls, and web servers.
Experimental Evaluation
The researchers evaluated FEEDAN using two prominent educational benchmarks:
- KDDcup2010: Data from Intelligent Tutoring Systems (ITS) containing 30 million rows, used to predict whether a student could solve a math problem on their first attempt.
- KDDcup2015: A dataset covering 39 MOOC courses, used to predict student dropout behaviors based on enrollment and learning logs.
Key Results
The experimental results demonstrate that privacy and performance are not a zero-sum game. FEEDAN proves that institutions can achieve "centralized-level" accuracy without the associated privacy risks.
- Performance Parity: In both dropout and performance prediction, FEEDAN's accuracy consistently approached the performance of a centralized model—the theoretical "gold standard" for accuracy.
- Superiority over Local Models: FEEDAN significantly outperformed models trained only on an individual institution's local data. This proves that "breaking the borders" of data islands provides a tangible boost in predictive power.
- Convergence and Scalability: While centralized training converges fastest, FEEDAN follows a similar trajectory, reaching high accuracy (often above 80-90%) within a few hundred epochs.
Conclusion and Future Directions
FEEDAN offers a robust pathway for the educational sector to leverage big data while maintaining strict student confidentiality. However, the authors identify three main hurdles for widespread adoption:
- Data Standards: The industry requires a unified data structure standard so models can easily identify relevant features across disparate databases.
- Public Model Design: There is a need for "public learning models" designed by pedagogical experts to address common needs across the entire education sector.
- Incentives and Alliances: Forming "data alliances" requires navigating the willingness of competing institutions to collaborate for a common pedagogical goal.
Summary updated