Abstract
Evaluating the quality of Massive Open Online Courses (MOOCs) is traditionally a manual, expert-driven process that lacks scalability. This paper introduces IaGRL (Information-aware Graph Representation Learning), a framework that treats MOOC quality evaluation as a graph representation learning problem. By modeling courses, students, and teachers within a Heterogeneous Information Network (HIN) and utilizing Mutual Information (MI) to ensure representation validity across multiple perspectives, the model significantly outperforms standard graph neural networks in predicting course quality.
Introduction
The rapid expansion of Massive Open Online Courses (MOOCs) has created a significant challenge for quality assurance. Traditional evaluation methods—manual audits by domain experts—are slow and expensive. Conversely, student-generated reviews are often subjective and biased. Automated evaluation offers a scalable alternative, yet existing methods often fail to capture the complex, multi-layered relationships inherent in educational ecosystems. A course is not an isolated entity; it is defined by its instructors, the students it attracts, and its thematic categorization.
Methodology
The researchers propose the Information-aware Graph Representation Learning (IaGRL) framework, which operates through two primary stages: multi-view representation and validity enforcement.
1. Multi-View Heterogeneous Information Network
The framework begins by constructing a MOOC Heterogeneous Information Network (HIN) consisting of four node types: Students (U), Teachers (T), Courses (C), and Subjects (S). To extract semantic meaning from this network, the authors utilize meta-paths—predefined sequences that capture specific relational views:
- Course-Teacher-Course (MP1): Captures courses linked by shared instructional expertise.
- Course-Student-Course (MP2): Captures courses linked by common student enrollment patterns.
- Course-Subject-Course (MP3): Captures courses within the same academic field.
Graph Convolutional Networks (GCNs) are employed within an encoder-decoder architecture to learn distinct embeddings for each of these views.
2. The Validity Guarantee (Information-Awareness)
To ensure that the learned mathematical embeddings are consistent with real-world data, IaGRL employs Mutual Information (MI) maximization. This "Information-Aware" component enforces three layers of validity:
- Raw Portfolio Agreement: Learned representations must remain faithful to raw course features (extracted from headlines and abstracts via Doc2Vec).
- Multi-View Consistency: Specific views (e.g., the teacher-centric view) must align with the final unified course representation.
- Course-Platform Alignment: Individual course embeddings are aligned with a global platform-level representation (the aggregate of all courses) to ensure contextual relevance.
Experimental Results
Dataset and Performance
The model was evaluated on a real-world MOOC dataset (2015–2018) comprising 10,919 courses, 4,931 students, 1,213 teachers, and 35 subjects. IaGRL demonstrated superior performance against several baseline models:
- Accuracy: IaGRL achieved 0.4101, nearly doubling the performance of standard GCN (0.2368) and outperforming GATv2 (0.3252).
- Macro-F1: The model reached 0.2011, indicating a robust ability to handle imbalanced data compared to other graph-based baselines.
Key Insights
- The Influence of Instruction: The "Teacher View" (MP1) emerged as the most critical factor, with an importance weight of 0.4655.
- Student Behavior Context: The "Student View" (MP2) had the lowest impact (0.2242). The authors hypothesize this is due to the dataset's focus on primary education, where student behavior is often teacher-guided rather than purely autonomous.
- Ablation Studies: Removing any of the Mutual Information loss functions (Raw Portfolio, Multi-View, or Platform Alignment) resulted in a performance drop, confirming that the "validity guarantee" is essential for high-quality evaluation.
Conclusion and Future Work
IaGRL represents a significant advancement in automated MOOC evaluation by integrating multi-view semantics with information-aware constraints. While highly effective, the authors note that the model's reliance on a rich HIN may present "cold start" challenges for new courses with limited interaction data. Future research will explore temporal dynamics to understand how course quality evolves as student-teacher interactions change over time.
Summary updated