Empowering education: Harnessing ensemble machine learning approach and ACO-DT classifier for early student academic performance prediction

Abstract

Higher education institutions have consistently strived to provide students with top-notch education. To achieve better outcomes, machine learning (ML) algorithms greatly simplify the prediction process. ML can be utilized by academicians to obtain insight into student data and mine data for forecasting the performance. In this paper, the authors proposed an ML-based student prediction model based on the demographic, social, psychological, and economic factors, collectively. The dataset utilized for this study was compiled from a designed questionnaire administered to second-year undergraduate students. The objective of this study is to uncover factors that could assist in predicting students' performance. Eight ML classifiers, logistic regression, random forest, support vector machine, XGBoost, support vector machine with a linear kernel, naïve Bayes, K-Nearest Neighbor, and decision tree are used to forecast student performance. Additionally, nine feature selection techniques, variance threshold, XGBoost, feature importance, recursive feature elimination, chi-square, ridge, Pearson correlation, lasso, and random forest, are employed to determine optimal factors. The authors experimented with each technique by creating two sets of training and testing data with 80:20 and 70:30 proportions, respectively. Comparatively, the ensemble DXK (DT + XGB + KNN) model with cross-validation and 80:20 proportions outperformed other standard classifiers, achieving a highest accuracy of 97.83%, an r-square of 96.17%, a precision of 97.94%, a recall of 97.83%, and an f1-score of 97.88%. These were the highest among all models tested. Additionally, the authors propose the ACO-DT model, which improves the prediction performance of the top-performing DT classifier by utilizing the Ant Colony Optimization technique. The findings demonstrate that the proposed model with 80:20 proportions achieve an accuracy of 98.15%, an f1-score of 98.16%, a precision of 98.18%, a recall of 98.15%, and an r-square of 84.75%, surpassing all other models for forecasting student performance. Using the specified data size, this model creation time is 8.49 s. The authors also recommended the future research directions to further enhance this study.