Uncertainty-Driven Student Performance Prediction Based on LSTM

With the rapid growth of online education platforms, predicting student performance has become a critical task for improving learning outcomes and personalizing education. In this study, we propose a novel framework for student performance prediction by integrating advanced machine learning techniques and uncertainty learning. Our approach combines Random Forest and Genetic Algorithms for feature selection, extracting both static and temporal features to capture the dynamic nature of student learning behaviors. We further enhance the model by incorporating Long Short-Term Memory (LSTM) networks with an attention mechanism and uncertainty-driven training, which focuses on difficult samples to improve robustness and predictive accuracy. Experimental results demonstrate the effectiveness of our framework, achieving significant improvements in F1-score, which reaches 0.861. The proposed framework not only supports personalized learning and early intervention strategies but also provides a foundation for future research in educational data mining and adaptive learning systems.