The Investigation of Influence Related to Feature Dimensionality Processing on Deep Learning-Based Student Performance Prediction

Predicting student performance is crucial for improving educational quality and personalized teaching. This study uses datasets from mathematics and Portuguese language courses to examine the impact of feature dimensionality processing on predictive models using an Artificial Neural Network (ANN) model. The study employs Principal Component Analysis (PCA) for dimensionality reduction and polynomial feature expansion to increase feature dimensions. The datasets were sourced from Kaggle, and PCA and polynomial feature expansion were applied to process feature dimensions. PCA was used to reduce dimensions, selecting 1 to 10 key dimensions to observe model performance under different dimensionalities. Polynomial feature expansion extended feature dimensions to capture nonlinear relationships, expanding up to the second to fifth degrees. The ANN model comprised an input layer, multiple hidden layers, and an output layer, utilizing Rectified Linear Unit (ReLU) activation functions and the Adam optimizer for training. Results indicate that the baseline ANN model exhibited higher prediction errors in mathematics, highlighting the complexity of predicting mathematical performance. PCA effectively reduced computational complexity and maintained high predictive performance when applied with appropriate dimensionality. However, polynomial feature expansion did not effectively improve model performance for either dataset and significantly increased errors for the mathematics dataset. This suggests that appropriate feature processing techniques are crucial for different datasets. This study recommends using PCA for dimensionality reduction in student performance prediction tasks and adjusting dimensionality based on dataset characteristics to optimize model performance and computational efficiency.