A classification study on predicting academic success using machine learning with oversampling and undersampling

Abstract

Machine learning represents a domain within artificial intelligence enabling computers to acquire the capacity for data analysis, model construction, and decision-making. Machine learning plays an important role in education. In this research, classifications were performed on the Open University Learning Analytics dataset using various machine learning techniques to predict students' academic performance. The success of various methods was evaluated, and to enhance the predictive accuracy of the classification algorithms data augmentation was performed using SMOTE, KMeansSMOTE, RandomOverSampler ADASYN, BorderlineSMOTE and SVMSMOTE oversampling techniques, which are frequently seen in the literature. Additionally, data reduction was performed using EditedNearestNeighbors, AllKNN, NearMiss, NeighborhoodCleaningRule, OneSidedSelection, RandomUnderSampler and TomekLinks undersampling techniques. Seven different machine learning techniques were used: Logistic Regression, Linear Discriminant Analysis, Random Forest, K-nearest neighbors, Decision Trees, Naive Bayes and Support Vector Machines. In addition to classification without resampling, ninety-one classification operations were performed using six different oversampling and seven different undersampling techniques. All classification processes are reported with four different performance metrics. Using the RandomOverSampler oversampling technique, 97% accuracy was achieved with the Random Forest classifier, and 96% accuracy was achieved with the K-nearest neighbors classifier using the AIIKNN undersampling technique. It has been noted that employing oversampling and undersampling techniques enhances the effectiveness of classifiers by addressing class imbalance within the dataset.