Enhancing Student Performance Management through Machine Learning and Deep Learning Models

Abstract

This work addresses the challenge of predicting student performance by investigating the sparsity of student action data and the imbalance between performance and completion rates. Initially, various machine learning (ML) and deep learning (DL) methods were compared using the Unitelma dataset, highlighting the superior performance of DL algorithms such as CNN and LSTM. Subsequently, an attention-based model combining convolutional and recursive layers (with LSTM units) was implemented and trained, although insufficient training data limited accurate model assessment. Lastly, the study focused on LSTM, TCN, and Kalman Filter models using the XuetangX dataset, leveraging oversampling (ADASYN) and data densification (PCA) techniques to address class imbalance and data sparsity issues. The Kalman Filter model demonstrated superior performance in terms of AUCPR, while LSTM and TCN models outperformed it in binary classification. TCN, in particular, showed increased efficiency over LSTM, especially for longer time sequences. Future work will involve applying these techniques to academic datasets, potentially retraining models with ADASYN and PCA algorithms to further improve performance. Additionally, emphasis will be placed on exploring the efficacy of TCN models in solving student performance prediction tasks.