AI-Powered Student Performance Prediction and Academic Risk Analytics in Higher Education: A Class-Decomposed XGBoost Framework

Abstract

Student dropout remains a significant challenge in higher education, resulting in substantial academic, financial, and institutional consequences. This paper presents a machine learning framework for early identification of students at risk of dropout, combining XGBoost with an engineered set of academicprogression ratio features and a Class Decomposition (CD) strategy: the majority (non-dropout) class is partitioned into k homogeneous sub-clusters via k-means prior to training, so the classifier is fit as a multi-class problem and the resulting sub-cluster probabilities are merged back into the original two classes at inference time. The proposed framework is evaluated on the publicly available Predict Students’ Dropout and Academic Success benchmark dataset under a binary dropout/non-dropout framing. At the default 0.5 decision threshold, the proposed XGBoost model reaches an AUC-ROC of 0.940 and an accuracy of 90.0% on a held-out test set. Hyperparameters are selected via randomized search optimizing AUC under stratified crossvalidation, and an alternative operating threshold can be separately calibrated using out-of-fold training predictions, avoiding test-set leakage. Comparisons with recent studies using the same benchmark dataset show that the proposed framework achieves the strongest reported AUC and accuracy among prior work restricted to this dataset and evaluation protocol, without relying on synthetic oversampling. Beyond raw predictive performance, we formalize the Class Decomposition procedure as a modelagnostic pre-training step (Algorithm 1), discuss deployment as a two-stage early-warning module inside institutional Student Information Systems (SIS), and outline the threats to validity and reproducibility considerations relevant to adopting the pipeline at a new institution. The findings suggest that the proposed pipeline is a practical, auditable foundation for institutional early-warning systems.