A Data-Driven Digital Twin for Student Engagement Prediction in e-Learning Systems

Machine Learning (ML) models are increasingly applied to Learning Management System (LMS) data to predict student engagement and performance. LMS data often contain missing values that can be informative. However, existing modeling approaches in education remove or impute missing values, which can lead to inaccurate or biased models. In this paper, we propose the use of digital twins to model students' engagement based on their learning activities on LMS while preserving the missingness patterns. We leveraged synthetic data generators such as Conditional Tabular Generative Adversarial Network (CTGAN), Tabular Variational Autoencoder (TVAE), and RealTabFormer with reversible data transformations to create a virtual replica of students' data. The CTGAN and TVAE generated balanced synthetic data that accurately captured the meaningful patterns of the real data. Moreover, XGBoost trained on a balanced virtual replica of the students' learning activities data obtained an F1-score of above 80% in predicting the students' engagement levels when evaluated on real data with both complete and incomplete entries. Our findings demonstrate how digital twins can be used to address the complexities of data in the education sector, improve the generalization of models, and reduce bias in real-world performance.