Performance assessment of machine learning models using SMOTE-ENN and feature selection techniques on student learning assessment criteria

Abstract

Imbalanced datasets, high-dimensional features, and the presence of redundant attributes often leads to poor classification performance and reduced model interpretability. This study addresses these limitations by proposing a hybrid methodology that integrates SMOTE-ENN with Information Gain, wrapper, and embedded feature selection technique. The aim is to enhance the prediction accuracy and improve the interpretability of classification models applied to student assessment data collected from tertiary institutions in Nigeria. The procedure involved training four classification models Random Forest (RF), Support Vector Machine (SVM), Logistic Regression (LR), and XGBoost on resampled datasets with selected features. The evaluation is done using 10-fold cross-validation and performance metrics which include accuracy, precision, recall, F1-score, and AUC. The results show that SVM achieved the highest performance across all feature selection methods, recording a peak accuracy of 0.96 and AUC of 0.99 when used with both Information Gain and Embedded RF features. RF performed robustly with a maximum accuracy of 0.95 and AUC of 0.98. XGBoost showed strong but slightly lower performance, while LR consistently recorded the lowest results, with a maximum accuracy of 0.71 and F1-score of 0.60.  The findings confirm that integrating hybrid feature selection with SMOTE-ENN significantly improves model performance, particularly for kernel-based and tree-based classifiers.