Multimodal Transformer-Based Framework for Predicting Student Engagement, Achievement and Creativity

The fast growth of online learning systems in India has created a variety of data that include academic, behavioural and linguistic aspects of student learning. Multimodal information integration is required to produce complex outcomes, including engagement, accomplishment, and creativity; however, traditional machine learning algorithms, such as decision trees and random forests, only prove their effectiveness when dealing with structured information. Two frameworks would be compared in this research paper: one of them the Transformer-based model (DistilBERT structured features), and the other one the Random Forest with a baseline where only engineered features are involved. A detailed prediction framework was created based on the creation of an over 23,000 student record of performance scores, screen-time trends and forum activities. The results of experiments indicate that the Transformer model achieves considerably greater scores than the Random Forest, to a lesser extent, in classification and regressive problems, particularly in creative evaluation by taking into account semantics and syntax as well as nuanced behavioural indicators of participation. In a move to make the analysis interpretable, SHAP-based analysis has highlighted the impact of the most important predictors, which include exam scores, time spent studying, screen time and lexical diversity. The research demonstrates the usefulness of multimodal AI in educational analytics in practice indicating the way forward to investigate scalable, explainable, and adaptive learning systems.

Перевод пока недоступен