AI - Powered Student Dropout Prediction Using AutoGluon FT- Transformer

Student retention and academic success are critical challenges in higher education, where institutions strive to iden-tify factors influencing student dropout, enrollment continuity, and graduation. In this study, we aim to predict student out-comes-Dropout, Enrolled, or Graduate-based on demographic, academic, and socioeconomic factors. A key challenge in this multi-class classification problem was the imbalance in class distribution, particularly for the Enrolled category, which re-sulted in lower recall and F1-score. To address this, we employed multiple machine learning models, evaluated their performance, and implemented data balancing techniques to enhance model effectiveness. We initially trained Random Forest and CatBoost, achieving 77.17% and 76.61 % accuracy, respectively. These models performed well for Dropout and Graduate predictions but struggled to accurately classify Enrolled students. We then leveraged AutoGluon's FT- Transformer, a deep learning model tailored for tabular data, which achieved 75.59 % accuracy but showed more balanced precision and recall across all classes. To overcome the class imbalance issue, we applied SMOTENC (Synthetic Minority Over-sampling Technique for Nominal and Continuous variables), generating synthetic samples for the underrepresented Enrolled class. This led to a significant performance boost, improving the FT- Transformer's accuracy to 78 %, surpassing all previously trained models. Our findings highlight the importance of addressing class imbalance in educational data analysis. By combining advanced machine learning models with targeted data augmentation techniques, we demonstrated that predictive performance can be significantly enhanced. This study underscores the potential of data-driven interventions in identifying at-risk students and improving educational strategies