Modeling students' performance using graph convolutional networks

Many models were recently proposed to classify students, relying on a large amount of pre-labeled data to verify their classification effectiveness. However, those models lack to accurately classify students into various behavioral patterns, employing nominal class labels, rather than ordinal ones. Meanwhile, such models cannot analyze high-dimensional learning behaviors among learners according to students' interaction with course videos. Since online learning data are huge, the main challenges associated with data are insufficient labeling and classification using nominal class labels. In this study, we proposed a model based on Graph Convolutional Network, as a semi-supervised classification task to classify students' engagement in various behavioral patterns. First, we proposed a label function to label datasets instead of manual labeling, in which input and output data are labeled for classification to provide a learning foundation for future data processing. Accordingly, we hypothesized four behavioral patterns, namely (High-engagement, Normal-engagement, At-risk, and Potential-At-risk) based on students' engagement with course videos and their performance on the assessments/quizzes conducted after. Then, we built a heterogeneous knowledge graph representing learners, course videos as entities, and capturing semantic relationships among students according to shared knowledge concepts in videos. Our model intrinsically works for heterogeneous knowledge graphs as a semi-supervised node classification task. It was evaluated on a real-world dataset across multiple settings to achieve a better predictive classification model. Experiment results showed that the proposed model can predict with an accuracy of 84% and an f1-score of 78% compared to baseline approaches.

Automated summary

This study proposes a model based on Graph Convolutional Network to classify students into different behavioral patterns based on their engagement with course videos. By introducing a label function to label datasets, the model provides a learning foundation for future data processing. Experimental results show that the proposed model achieves high predictive accuracy and f1-score.