Federated Learning Approaches for Privacy-Preserving Educational Data Mining

The recent explosion in educational data produced by digital learning systems, learning management systems, and intelligent tutoring systems have presented tremendous opportunities to educational data mining (EDM) to improve learning outcomes and personalize instruction and inform decisions based on data. The core element of the centralized gathering and evaluation of sensitive student data, however, provoke genuine security, ethical, and legal issues especially when considering information protection regulations like GDPR and FERPA. Federated Learning (FL) has become an attractive paradigm of privacy-preserving educational data mining that learners can train models collaboratively across distributed educational institutions without having to transfer raw data to a central server. This paper is an in-depth analysis of federated learning methods in educational data mining, in terms of optimization of their data mining capabilities with the analytical performance, data privacy and security implications. The paper describes the basic FL architectures, such as horizontal, vertical and hybrid federated learning, and qualifies them to suit the typical EDM tasks (student performance prediction, dropout detection, learning behavior analysis and recommendation system). The main issues like data heterogeneity, communication efficiency, model convergence, and risks of privacy leaks are examined severely in the context of the educational settings. In addition, the paper also examines recent additions to federated learning, such as secure aggregation, differential privacy, and personalized federated models, and how they can enhance privacy guarantees without affecting model accuracy. This paper, by synthesising the existing knowledge and pointing at the future directions, will show that federated learning is a viable and scalable privacy preserving educational data mining, which will facilitate further growth of educational analytics without violating the rights of the learners as the data belongs to them.