AutoML Algorithms for Predicting Students’ Dropout and Academic Success in Higher Education

Abstract

Challenges with high dropout rates and low academic achievement continue to be a major issue for higher education institutions. To deal with these challenges, it is essential to identify the underlying causes and develop strategic solutions to support student success. Machine learning techniques are powerful tools for predicting student academic performance and enabling timely, targeted interventions. In this study, the goal is to test Automated Machine Learning (AutoML) frameworks that can predict student dropout and academic success. The dataset included students' demographic, socio-economic, and academic information gathered from institutional records. H2O AutoML, TPOT, PyCaret, and AutoGluon were evaluated, as these tools automate important parts of the machine learning pipeline, making it easier to train and select models. Evaluation metrics including accuracy, precision, recall, and F1-score, as well as training time and CPU usage, were used to assess model performance and computational efficiency. Three main experiments were conducted: the first evaluated model diversity, the second observed the effect of cross-validation fold size, and the third assessed different search space configurations. In the first experiment, TPOT obtained the highest accuracy of 90.04%. In the second experiment, using 10-fold cross-validation, H2O AutoML obtained the best accuracy of 91.38%. In the third experiment, TPOT maintained its accuracy at 90.04% across all configurations. A comparison between the conventional ExtraTreesClassifier and TPOT's AutoML ExtraTreesClassifier showed comparable performance, with the conventional model achieving slightly higher accuracy of 90.20% and recall of 94.44%, while TPOT demonstrated superior precision of 90.24%. These results demonstrate that AutoML frameworks are effective tools for early identification of at-risk students, offering competitive performance with significantly reduced manual effort. Future studies could explore deploying these models within academic advising systems.