Modeling students' performance using graph convolutional networks.

Area of Science:

• Educational Technology
• Computer Science
• Machine Learning

Background:

• Traditional student classification models struggle with high-dimensional learning behaviors and nominal labels.
• Insufficient data labeling is a major challenge in analyzing large online learning datasets.
• Existing methods fail to capture nuanced student engagement patterns from video interactions.

Purpose of the Study:

• To propose a Graph Convolutional Network (GCN) based semi-supervised classification model for student engagement.
• To develop an automated labeling function to overcome manual labeling limitations.
• To classify students into four distinct behavioral patterns: High-engagement, Normal-engagement, At-risk, and Potential-At-risk.

Main Methods:

• Constructed a heterogeneous knowledge graph representing learners and course videos as entities.
• Utilized a semi-supervised node classification task on the knowledge graph.
• Developed a novel label function for automated dataset labeling.
• Hypothesized four student behavioral patterns based on video engagement and assessment performance.

Main Results:

• The proposed GCN model achieved 84% prediction accuracy.
• The model demonstrated a 78% f1-score, outperforming baseline approaches.
• The semi-supervised approach effectively classified students into hypothesized behavioral patterns.

Conclusions:

• The novel GCN model offers an effective solution for classifying student engagement in online learning.
• Automated labeling and heterogeneous knowledge graphs enhance classification accuracy.
• The model provides a robust framework for analyzing complex learning behaviors.