Predicting Student Academic Performance Using Meta-Learning with Model-Agnostic Meta-Learning (MAML)

Abstract

Prediction of student academic performance is one of the important elements of educational analytics since it enables early interventions, customized learning experiences, and decision-making processes at educational institutions. Unfortunately, classical approaches of machine learning and deep learning algorithms show their low adaptability in diverse academic environments with different curricula, behaviors, and learning styles of students. Furthermore, some educational institutions suffer from a lack of annotated data, and, therefore, traditional prediction systems fail to provide accurate results. This study introduces a novel approach of meta-learning based on Model-Agnostic Meta-Learning (MAML) for predicting the academic performance of students adaptively. The approach leverages several educational datasets as meta-learning tasks and trains the model to obtain initial weights that can quickly adapt to new academic environments using only a few examples. This model combines elements from academic, behavioral, demographic, and online participation factors to form the deep neural network prediction model. An experiment has been performed on standard educational data sets, and the results have been compared with other baseline models such as Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), XGBoost, Artificial Neural Network (ANN), and Long Short-Term Memory (LSTM). From the experimental results, it is clear that the proposed MAML model performed better than any other model, with a classification accuracy of 95.3% as compared to LSTM (92.4%) and XGBoost (90.2%). The proposed model gave an MAE of 0.24, an RMSE of 0.35, and an R² of 0.95 for regression-based GPA prediction, demonstrating its high prediction reliability and consistency. This model also showed good few-shot learning ability and cross-domain generalization ability in the low-data educational domain. Thus, it can be concluded that meta-learning can greatly improve the adaptability, scalability, and efficiency of intelligent educational prediction systems.