Abstract
As Massive Open Online Courses (MOOCs) become increasingly popular, providing extracurricular knowledge for users is a promising direction. However, existing methods struggle with semantic drift and a lack of knowledge guidance in complex MOOC environments. This paper introduces a novel three-stage framework that establishes a dynamic boundary for concept searching via external knowledge bases and utilizes heterogeneous features for verification. Furthermore, an interactive optimization mechanism based on a game involves human effort in the model's refinement. Experiments on datasets from Coursera and XuetangX demonstrate that this method achieves significant improvements (+0.19 MAP) over existing state-of-the-art methods.
Introduction: The Challenge of Concept Expansion
Massive Open Online Courses (MOOCs) are often constrained by time, meaning instructors may introduce a core concept—such as a Binary Search Tree—without the opportunity to explore related extracurricular topics like Heaps, Priority Queues, or Tango Trees.
Automatically identifying these relevant additions is known as "Course Concept Expansion" (CCE). Traditional set-expansion algorithms fail in this technical domain due to two primary issues: semantic drift (e.g., drifting from computer science "Trees" to botanical "Trees") and the complexity of heterogeneous relationships between concepts, which may be prerequisites rather than simple synonyms.
Methodology: A Three-Stage Framework
The researchers proposed the MOOC-CG model to address these challenges through a structured pipeline:
1. Candidate Generation (The Boundary)
To prevent semantic drift, the model treats the course as a "concept space" made of clusters. It uses an unsupervised embedding-based algorithm to build a dynamic boundary within an external Knowledge Base (XLORE). If a potential new concept is too distant from the cluster center, it is rejected immediately.
2. Concept Classification (The Filter)
Once candidates are identified, an XGBoost classifier determines their quality based on three key features:
- Confidence Score: Proximity to the concept cluster.
- Search Path Encoding: An RNN encodes the specific path taken through the Knowledge Base (e.g., BFS → InstanceOf → Graph Algorithms). This helps the model understand the nature of the relationship.
- Prerequisite Features: Evaluating if the concept is a necessary foundation for the current course material.
3. Game-based Optimization (Human-in-the-Loop)
To refine the model with human intelligence, the authors deployed a game called "Top-Student" on the XuetangX platform.
Students earn points and improve their rank (e.g., reaching "Silver Level") by "deleting" low-quality or irrelevant expanded concepts. This crowdsourced feedback provides high-quality labels used to iteratively retrain the model. A group-vote scoring mechanism prevents irresponsible clicking by penalizing users who delete concepts that others find helpful.
Key Results and Insights
The MOOC-CG framework was tested on four datasets across two domains (Data Structure and Algorithm, and Psychology) and two languages (Chinese and English).
- Significant Performance Gains: The proposed method achieved an average improvement of +0.19 in Mean Average Precision (MAP) over existing state-of-the-art methods.
- Importance of Paths: Ablation studies showed that "Search Path Encoding" was the most critical feature; removing it caused the F1-score to drop by nearly 8%. Knowing how concepts are connected is more important than simple proximity.
- Stability: While baseline methods fluctuated between subjects, this framework remained stable across both technical (DSA) and social science (PSY) domains.
Conclusion
By combining external knowledge graphs with gamified human feedback, this research offers a robust solution for enriching MOOC curricula. The success of the "Top-Student" game suggests that integrating data labeling into the learning process can effectively bridge the gap between automated algorithms and human expertise.
Summary updated