What question did this study set out to answer?

This research aims to develop a cohesive machine learning framework that enhances predictive school management through data analysis.

April 6, 2026Open Access

A machine learning based framework for predictive school management using student and faculty analytics

Key Points

This research aims to develop a cohesive machine learning framework that enhances predictive school management through data analysis.
Developed an integrated machine learning system using various AI technologies.
Tested the framework on UCI benchmark datasets with strict nested cross-validation.
Utilized deep learning methods (LSTM, GRU, CNN, Transformers) for academic growth modeling.
Implemented gradient boosting ensembles (XGBoost, LightGBM, CatBoost) for data analysis.
Employed graph convolutional networks to understand relationships between students, faculty, and courses.
Achieved predictive accuracy of 99.6% and 97.5% for at-risk student classification.
Displayed high recall rates of 99.4% and 96.7% in identifying at-risk students.
Demonstrated over a 5% improvement compared to top baseline models.
Performed well under conditions of feature noise and missing data, maintaining accuracy above 96%.
Minimized gender and age bias, and improved fairness metrics for parental education gap.

Abstract

New technologies in education have created a huge amount of data that, when used effectively, can have a major impact on the functioning of an institution and the academic achievement of students. Nevertheless, all existing predictive models are still disconnected and do not integrate historical trends, student-faculty relationships, and trend patterns into a coherent decision-making system. The paper describes an integrated machine learning system that integrates several synergistic AI technologies: (1) deep learning systems (LSTM, GRU, CNN, and Transformers) to model academic growth over time; (2) comprehensible gradient boosting ensembles (XGBoost, LightGBM, and CatBoost) to understandably infer and analyze structured data. (3) graph convolutional networks (GCNs) to encode academic relationships between students, professors, and courses; and (4) data-centric oriented approaches (multitasking, transfer, and federated learning). The framework is tested on two UCI benchmark datasets (n = 649) with fully isolated holdout sets using strict nested cross-validation to prevent data leakage. The framework yields 99.6% and 97.5% predictive accuracy (5.6% and 6.3% improvement over the top baselines) and high recall (99.4% and 96.7%) in classifying at-risk students. Each component has been shown to contribute fully in ablation studies, and the hybrid framework has been shown to outperform state-of-the-art transformed table models (TabTransformer, FT-Transformer, and SAINT) (99.6% vs. 97.2% for the best transformer). Robustness analysis with feature noise and missing data (> 96% accuracy with 20% missing data) demonstrates excellent regression. Fairness assessment indicates that gender and age bias are very small, and mitigation strategies (reweighting, adversarial debiasing) bring the parental education gap down to 0.1%. Cross-domain experiments (mathematics/Portuguese) show a performance loss of -2.3%, indicating internal generalizability, but cross-institutional validation remains to be performed. This framework provides educators with interpretable, actionable insights into evidence-based interventions, demonstrating that for accurate, fair, and robust predictive educational analytics, multi-paradigm AI integration is essential and comprehensive.

Mark Helpful

Bookmark

Relay

View Full Paper