A machine learning framework with a public synthetic data set for early student profiling and outcome prediction

The accurate and early prediction of student academic outcomes is fundamental to enabling timely interventions and personalized support in higher education. However, progress in educational data mining is severely constrained by the scarcity of publicly available data sets, as student records are protected by stringent privacy regulations that prohibit their distribution. This paper introduces a comprehensive machine learning framework for early student profiling and outcome prediction, underpinned by a novel public synthetic data set of over 100,000 student records designed specifically for machine learning research in educational contexts. The data set encompasses 28 attributes spanning demographic characteristics, academic background, entrance examination scores, socio economic indicators, and behavioral metrics, incorporating realistic complexities including non linear relationships, systematically introduced missing value patterns, and controlled class imbalance. Two complementary prediction tasks are defined: ordinal classification of student academic level into Beginner, Intermediate, Advanced, and Exceptional categories, and nominal classification of division allotment into Remedial, Regular, Advanced, and Honors divisions. A modular machine learning framework is developed encompassing comprehensive preprocessing, feature engineering including composite indices and interaction terms, implementation of eleven distinct algorithms ranging from interpretable baselines to ensemble methods and neural networks, rigorous cross validation, and multi faceted evaluation incorporating accuracy, per class metrics, macro averaged F1 scores, and ordinal specific measures. Experimental results demonstrate that ensemble methods substantially outperform simpler approaches, with LightGBM achieving macro F1 scores of 0.842 for student level prediction and 0.826 for division allotment. An ordinal neural network incorporating the ordinal nature of the student level target achieves the highest overall performance with macro F1 score of 0.846 and quadratic weighted kappa of 0.892, confirming the value of explicit ordinal modeling. Feature importance analysis reveals that prior academic achievement, particularly class 12 percentage, dominates predictions, while socio economic and behavioral factors provide meaningful secondary contributions. SHAP analysis uncovers important interaction effects, including the moderating role of prior achievement on the impact of entrance examination scores, and provides local explanations that enable targeted intervention design. Cross validation confirms stability with standard deviations below 0.011 across folds, and computational efficiency analysis identifies LightGBM as offering optimal balance between predictive performance and resource requirements. The complete framework, including all preprocessing modules, model implementations, evaluation protocols, and interpretation tools, is released as open source software alongside the validated synthetic data set, providing the research community with foundational resources to advance reproducible educational data mining research without privacy constraints. This work establishes a comprehensive benchmark for student outcome prediction and provides actionable insights for developing equitable, data driven student support systems.