Dimensionality Reduction using Feature Selection Techniques on EDM for Student Academic Performance Prediction

Abstract

The aim of this study is to perform dimensionality reduction on student datasets using different feature selection strategies in order to better forecast students' academic progress and minimize high-dimensionality concerns. The effectiveness of models built utilizing classification algorithms and pertinent features chosen using different feature selection strategies has been experimentally evaluated in this study. Nine feature selection methods, including Chi-Square, Pearson Correlation, Variance Threshold, Feature Importance, Recursive Feature Elimination (RFE), Lasso, Ridge, Random Forest, and XGBoost, as well as one classification method, logistic regression, were applied to the student dataset to determine the results of the analysis.  In this experiment, dimensionality reduction techniques were used to reduce the number of features from 70 to the optimal 21 feature subsets for analysis of student academic performance prediction. Five measures were used to assess the effectiveness of the techniques: r², precision, recall, f-score, and accuracy. With the logistic regression classifier, the feature subset chosen by the chi-square had the highest accuracy of 72.4%, precision of 71.4%, recall of 72.4%, f-score of 70%, and r-square of -0.221%.