Artificial Intelligence-Driven Academic Performance Early Warning System for Engineering Universities: Application of XGBoost-SHAP Algorithm

Le résumé fourni par la source

Prior research concerning academic warning and performance prediction for university students within the contemporary credit system primarily utilizes statistical or machine learning techniques. Nevertheless, prediction models frequently encounter issues related to the excessive incorporation of various attributes, which culminates in intricate models and challenges in acquiring data time series. This investigation leverages students’ historical academic records, employing machine learning interpretable methods to construct a prediction model. In dataset processing, an innovative application of hierarchical clustering was implemented, In this study, the HAC algorithm is compared with the DBSCAN algorithm, and the Silhouette score of the HAC algorithm is improved by 0.261; Calinski–Harabasz score increased by 154.615; Davies–Bouldin score decreased by 0.308, showing good model superiority. Therefore, students’ actual academic conditions are considered to categorize regular graduating students. Subsequently, a feature fusion method was applied to extract dataset features. Ultimately, the XGBoost model was trained and evaluated, and the SHAP interpretable model was employed for the explanatory analysis of the developed model. Experimental outcomes demonstrate that, in comparison to Random Forest, Lasso, SVR, LightGBM, and Catboost machine learning algorithm models, the accuracy of XGboost prediction model is improved by 8.4%, 22.6%, 20%, 1.5%, and 5.9%, respectively. This underscores its efficacy in predicting a student’s risk of extending their studies. The merits of this approach not only enhance the accuracy and interpretability of academic warnings but also propose a novel paradigm and trajectory for devising talent training programs in new credit-based engineering universities.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.