Explainable AI for Early Detection of Student Learning Difficulties

Abstract

In this study, a machine learning-based predictive analysis was carried out to identify a set of university students who may suffer from learning difficulties and the predictions were made interpretable using explainable AI tools instead of leaving them as an opaque, black-box output from the machine learning model. The data were gathered from 1200 students of universities, who were observed regarding attendance, assignment scores, quiz performance, examination results, learning-platform activity, and study hours. The data set was pre-processed by discarding of incomplete data sets and filling in missing values with the median, and relevant features were selected through recursive feature elimination and mutual information. The cleaned dataset was divided into 70:30 training and testing sets and three different classifiers were trained and evaluated using accuracy, precision, recall, F1-score, and ROC-AUC. Then, SHAP (SHapley Additive exPlanations) was used to determine the relative importance of the features in the best model to make individual predictions. The results showed that XGBoost was the best overall model using the five metrics followed by Random Forest and then SVM. The results of the feature selection methods based on mutual information and recursive feature elimination were similar, with a consistent core set of predictive features being the learning-platform activity, quiz performance and attendance. These findings were corroborated and expanded in the SHAP analysis of the XGBoost model, which identified low learning-platform engagement, falling quiz scores, and attendance as the top individual indicators of at-risk status, with the lowest level of learning-platform engagement having the largest effect. The study results suggest that using highly effective ensemble models with SHAP explanations is a viable, interpretable method for identifying at-risk students early, and provides guidelines for institutional implementation, ethical considerations, and future research.