Federated deep learning for early dropout prediction in online courses

Abstract

High dropout rates remain a persistent challenge in online education, undermining learner success, institutional credibility, and resource efficiency. Traditional dropout prediction approaches rely on centralized data collection, raising serious concerns regarding privacy, security, and compliance with data protection regulations such as GDPR and FERPA. This study proposes a Federated Deep Learning (FDL) framework that enables collaborative model training across multiple institutions without sharing raw learner data. The hybrid architecture combines 1D Convolutional Neural Networks (CNN) for spatial feature extraction with Bidirectional Long Short-Term Memory (Bi-LSTM) networks for temporal sequence modeling. An entropy-weighted aggregation strategy ensures balanced contributions from diverse datasets, while gradient compression optimizes communication efficiency. Using datasets from three heterogeneous institutions a MOOC platform, a university LMS, and a corporate e-learning environment the proposed model achieved an average F1-score of 0.89 and identified at-risk learners up to two weeks before dropout events. Comparative experiments against centralized deep learning and traditional machine learning baselines demonstrated superior accuracy, earlier detection, and enhanced generalization across contexts. The results affirm FDL as a technically robust, privacy-preserving, and ethically aligned approach for early dropout prediction, paving the way for scalable, secure, and inclusive educational analytics.