predicting students’ academic performance is one of the critical and challenging topics in the field of machine learning and higher education, as it plays a key role in identifying at-risk students and providing supportive and preventive interventions. While previous studies have mainly focused on traditional machine learning algorithms and data-driven techniques such as SMOTE, the present study adopts a more comprehensive and multidimensional approach.

The data used in this research were collected from multiple institutions and various courses to reduce the generalizability limitations of prior studies and to ensure that the findings can be applied across diverse educational settings. In addition to classical algorithms such as J48, Random Forest (RF), Support Vector Machine (SVM), k-Nearest Neighbors (kNN), Logistic Regression (LR), and Naive Bayes (NB), this study examines deep learning algorithms including Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), as well as advanced hybrid models such as XGBoost and AutoML.

To address the issue of imbalanced datasets and uneven distribution of training samples, in addition to SMOTE, ADASYN and cost-sensitive learning methods were employed to enhance the models’ ability to detect minority and critical samples. The results indicate that deep learning models are capable of capturing complex nonlinear relationships among features and students’ academic performance, while the XGBoost hybrid model achieves the best balance across Accuracy, Precision, Recall, and F-measure metrics.