A Multi-Algorithm Fusion Approach to Predicting Student Performance

Abstract

In contemporary blended learning environments that seamlessly integrate online and offline educational components, the ability to accurately predict students' academic performance using formative learning data has emerged as a critical factor in improving overall teaching quality. However, existing approaches often face significant challenges, particularly regarding the low accuracy and poor generalization capabilities typically associated with single prediction algorithms in student performance forecasting. To systematically address these persistent issues, this study constructs a robust multi-algorithm fusion model for performance prediction based on the Stacking ensemble learning framework. Specifically, this advanced model employs Random Forest, XGBoost, and LightGBM as foundational base learners to capture diverse data patterns, while utilizing Ridge Regression as the meta-learner to prevent overfitting. This hierarchical architecture achieves highly accurate course performance predictions through comprehensive secondary learning across multiple heterogeneous models. To rigorously validate the proposed methodology, the study utilized comprehensive grading data from the 'C Language Programming' course—taught simultaneously through online and offline modalities to computer-related majors at Hulunbuir University—as the primary experimental dataset. The comprehensive experimental results conclusively demonstrate that the proposed multi-algorithm fusion model significantly outperforms traditional single algorithms in terms of overall regression accuracy and predictive stability. Consequently, this innovative approach provides highly reliable predictive support and actionable insights for universities, enabling administrators and educators to conduct precise, data-driven teaching evaluations and implement proactive academic quality monitoring systems effectively.