A personalized federated deep learning framework predicts student performance in large-scale examinations across heterogeneous regional distributions

The predictive modeling of student achievement in national education systems faces a persistent conflict between the need for large-scale data utility and the stringent requirements of student data privacy. This study addresses this challenge through a high-performance personalized federated learning framework that predicts the General Success Score Percentiles of 1,094,954 students in a national high school entrance examination under a simulated regional federation in which model updates rather than raw regional records were exchanged during federated training. Utilizing a large dataset of 31 diverse predictors encompassing academic history, instructional quality, and socio-economic factors, the research compared three neural architectures consisting of a Multilayer Perceptron, a Federated Attention Model, and a Deep and Cross Network. To manage the inherent regional heterogeneity and non-independent and non-identically distributed data across the seven geographical regions, the study utilized the FedProx optimization algorithm. Results showed that the Personalized Federated Multilayer Perceptron achieved the highest overall predictive performance ($${R}^{2}=0.7972$$), modestly exceeding the centralized XGBoost ($${R}^{2}=0.7947$$) and centralized MLP ($${R}^{2}=0.7910$$) baselines while retaining the decentralized and region-adaptive advantages of federated learning without pooling raw regional records. The integration of Regional SHapley Additive exPlanations analysis provided a transparent mapping of feature importance, revealing that while prior academic performance is the primary national driver, environmental factors exert disproportionate influence in specific metropolitan and other regions. The findings indicate that regionally adaptive models can support targeted resource allocation, while cautioning that systems predicting attainment from environmental context risk formalising the disadvantage they measure. The predictive modeling of student achievement in national education systems faces a persistent conflict between the need for large-scale data utility and the stringent requirements of student data privacy. This study addresses this challenge through... [2600 chars]