A multimodal learning analytics model based on deep learning for predicting student performance using tabular and time-series data fusion

The rapid growth of digital learning has increased interest in using multimodal learning analytics (MLA) to better understand student behavior and support more personalized teaching. This study presents a deep learning based MLA model that leverages different types of educational data such as demographics, assessment scores, online activity, and learning style information through separate feature extraction branches and an intermediate data fusion method. Our model combines Multi-Layer Perceptron (MLP) and Long Short-Term Memory (LSTM) neural networks to automatically extract and learn features from both tabular and temporal data. Furthermore, an Aggregation block is also introduced to convert sparse intermediate features into dense and more useful representations, helping the model capture stronger relationships across modalities. The proposed approach enables early identification of student performance trends, thereby supporting more informed and individualized timely interventions. Our experiment on the Open University Learning Analytics Dataset (OULAD) and a self-collected dataset show that the proposed model archives promising results. On OULAD, it reaches an average Mean Squared Error (MSE) of 0.2404, Mean Absolute Error (MAE) of 0.3413 and 0.7836 R2-score value. Moreover, it also achieved competitive results on our self-collected dataset where it reaches 0.7410, 0.6636 and 0.2893 for average MSE, MAE and R2-score value respectively. These results demonstrate the value of combining multiple data sources and the benefits of the proposed architecture. Overall, this work offers a reliable and scalable approach to learning analytics that can help create more adaptive and effective digital learning environments.