Evaluating Privacy-Conscious Federated Learning Strategies for Reliable Student Performance Prediction in Learning Analytics

Federated learning can support privacy-conscious student performance prediction by enabling collaborative model training across educational institutions without centralizing sensitive student records. However, existing educational federated learning research is dominated by FedAvg, while alternative strategies such as FedProx are mainly explored in studies explicitly addressing non-independent and identically distributed (non-IID) educational data. This study addresses this gap through a controlled comparison of seven federated learning strategies, including FedAvg, FedAvgM, FedProx, FedAdagrad, FedAdam, FedYogi, and FedMedian, using the Open University Learning Analytics Dataset. A Flower-based simulated federated learning framework was developed to evaluate these approaches across multiple student outcome prediction tasks. For Pass/Withdrawn prediction, FedMedian achieved the highest mean late-round accuracy and macro-$F_1$. Hochberg corrected p-value showed significant differences between FedMedian and all six alternative methods for mean accuracy. For Pass/Fail prediction, FedMedian had the highest observed mean performance, but its advantage over the leading alternatives FedProx, FedAvgM, and FedAvg was not statistically confirmed under the corrected paired comparisons. FedProx had the highest observed late-round means for Fail/Distinction prediction but was not statistically distinguishable from FedAvg, FedAvgM, or FedAdagrad after Hochberg correction. None of the evaluated strategies achieved satisfactory balanced performance for the highly imbalanced Pass/Distinction prediction task, demonstrating the challenges of applying federated learning to complex educational outcomes. These findings indicate that FedAvg, despite being the most commonly adopted approach in educational federated learning, is not consistently optimal under heterogeneous educational data distributions. Instead, selecting federated optimization strategies according to task characteristics can improve model stability and prediction reliability. The study highlights the importance of responsible federated learning deployment in educational settings, where privacy-conscious analytics can support more reliable identification of student learning outcomes and inform evidence-based interventions.