Corresponding Author: Shubham Gupta
Given their impact on student achievement, efficiency in the college, and retention, student performance prediction and dropout detection have emerged as important areas of study in the field. Traditional approaches to identifying academically at-risk students are generally manual assessment and retrospective analysis, which are non-predictive and time-consuming. In order to overcome these challenges, this paper suggests using machine learning techniques to develop a predictive analytics framework for student performance and dropouts that is based on XAI. The proposed methodology includes data preprocessing, feature selection, predictive modeling using XGBoost, and explainability using SHAP to accurately predict students' academic performance with the aim of understanding the model's predictions. The data set used in this research is the UCI data set comprised of the demographic, academic, financial, and institutional data of 4,424 students. The various ML algorithms such as DT, SVM, RF, and XGBoost were used and compared using the various performance measures. The experimental outcomes showed that the proposed XGBoost model achieved higher accuracy (94.18%), precision (93.74%), recall (93.21%), F1 score (93.47%), and ROC-AUC score (95.62%). Further, the curricular unit performance, admission grades, and the student's status as a tuition fee-paid student were the most relevant for student academic outcomes and dropout, as identified through the SHAP-based explainability analysis. The proposed framework allows for the identification of the pupils who are at risk of underachievement from an early age, and to make informed decisions based on clear and understandable analytics. The results of this research suggest that the proposed system may have a significant advantage in educational institutions to help decrease the dropout rate, improve academic support, and better student retention by implementing proactive intervention systems.