Explainable Machine Learning for Student Performance Prediction in Higher Education
Authors-Shilpa K
Keyword-Academic Risk Prediction; Learning Analytics; XGBoost; Explainable Artificial Intelligence; Student Performance; Higher Education
In higher education the early identification of students who are at risk of academic underperformance is crucial to deliver timely and targeted academic support. In this study, an Adaptive Multi-Modal Learning Analytics Framework (AMLAF) is proposed that combines academic, behavioural, institutional, learning preference and perceptual variables for predicting student academic risk. Explanability analysis with SHAP and LIME revealed that previous academic performance, study behavior, mentoring, institutional support and engagement of students were relevant factors that influence the model predictions. Results showed that the multi modal features engineered using these features were used to train Random Forest, XGBoost and LSTM models, with the latter model proving to be the most accurate. XGBoost with test accuracy of 99.47%, AUC-ROC of 0.9992 and 99.59% mean cross validation accuracy was the best performing model that was observed. The model got 7 false negative predictions and 1 false positive prediction in the test set, and had a recall of 98% and an F1 score of 99% for the at-risk class. The simulated XGBoost–LSTM ensemble obtained an accuracy of 96.75%, which is the same as XGBoost to show no performance improvement over XGBoost alone. LSTM obtained an accuracy of 81.06%, while the simulated XGBoost–LSTM ensemble obtained an accuracy of 96.75%, showing no improvement over XGBoost alone. The results show the first application to supporting early identification of academic risk and evidence based student support planning through multi modal learning analytics, of which explainable is one.
Doi-[https://doi.org/10.5281/zenodo.23160425]