Predicting student grades via adaptive multi-level learning models.

Area of Science:

• Educational Data Mining
• Machine Learning in Education
• Higher Education Analytics

Background:

• Educational institutions utilize intelligent systems for student data analysis.
• Early prediction of student performance is crucial for academic advising and interventions.
• Existing methods struggle with temporal data sparsity in diverse educational settings.

Purpose of the Study:

• To propose an adaptive multi-level prediction framework for student course performance.
• To develop a model-agnostic system capable of handling data sparsity.
• To demonstrate the framework's generalizability and effectiveness across different institutional data.

Main Methods:

• Segmenting student-course data into homogeneous groups.
• Assigning temporally validated specialist models to each group.
• Implementing an automated data-density fallback guard for data sparsity.
• Utilizing a model-agnostic approach accepting various machine learning algorithms.

Main Results:

• Achieved an 18.6% RMSE improvement on a dense university dataset, favoring tree-based ensembles.
• On a sparse course dataset, the framework selected diverse models, including neural networks (48% of clusters) and collaborative filtering (22% of windows).
• Demonstrated strong cross-institutional generalizability on large-scale, real-world datasets.

Conclusions:

• The adaptive framework effectively shifts model configuration from manual heuristics to data-driven optimization.
• The system successfully adapts to varying data densities and curriculum structures.
• Results indicate significant improvements in student performance prediction accuracy and model selection efficiency.