Comparative Evaluation of Feature Selection Algorithms for Predictive Modeling of Academic Performance Outcomes

This paper explores the application of various feature selection methods in predictive modeling for academic performance, with a focus on machine learning techniques which is Multi-Layer Perceptron (MLP). By employing methods such as RFEMutual Information, Random Forest Importance, XGBoost Importance, and L1 Regularization Lasso (L1 Lasso), the study aims to identify the most relevant features from a dataset comprising 12 features related to student online engagement. The models were evaluated using accuracy, precision, recall, and F1-score to assess their comparative effectiveness across feature selection methods. The results highlight the effectiveness of RFE Ranking and L1 Lasso in achieving balanced model performance while reducing dimensionality, with these methods showing superior results compared to others in terms of generalizability and robustness. This study contributes to the ongoing discourse on educational data mining by offering insights into the practical implications of feature selection in enhancing predictive accuracy and guiding data-driven decision-making in educational settings.