Abstract
As Massive Open Online Courses (MOOCs) become increasingly popular, it is promising to automatically provide extracurricular knowledge for MOOC users. Suffering from semantic drifts and lack of knowledge guidance, existing methods can not effectively expand course concepts in complex MOOC environments. In this paper, we first build a novel boundary during searching for new concepts via external knowledge base and then utilize heterogeneous features to verify the high-quality results. In addition, to involve human efforts in our model, we design an interactive optimization mechanism based on a game. Our experiments on the four datasets from Coursera and XuetangX show that the proposed method achieves significant improvements (+0.19 by MAP) over existing methods.
Introduction
Massive Open Online Courses (MOOCs) often face a "Goldilocks" problem: instructors must keep courses concise for a global audience, which frequently results in the exclusion of helpful auxiliary concepts. For instance, a student learning about Binary Search Trees might benefit from knowing about Heaps or Tango Trees, even if they are not explicitly part of the syllabus.
Automatically suggesting these "out-of-teaching" concepts is a challenging task known as set expansion. Existing methods struggle in the MOOC environment due to two primary issues:
- Semantic Drift: In complex fields like Computer Science, models can easily drift from specific topics (e.g., data structures) into unrelated domains (e.g., general software engineering).
- Heterogeneity: Course concepts are diverse, encompassing structures, algorithms, and theoretical applications, making them harder to categorize than simple lists.
Methodology
The researchers developed a three-stage framework to address these challenges by combining external knowledge bases with human-in-the-loop optimization.
1. Candidate Generation with Dynamic Boundaries
Instead of an unconstrained search, the model utilizes an external Knowledge Base (XLORE). It treats the course as a "concept space" composed of multiple clusters. To prevent semantic drift, the model constructs a "hypersphere" (a multi-dimensional boundary) around these clusters. New concepts are only considered candidates if they fall within this boundary, tethering the expansion to the course's core themes.
2. Concept Classification
Candidates are filtered using an XGBoost classifier trained on three key features:
- Confidence Score: Proximity of the concept to the cluster center.
- Search Path Encoding: An RNN-based encoder that analyzes the structural relationship (e.g., sibling vs. descendant) within the knowledge graph.
- Prerequisite Features: Analysis of course text to determine if the concept serves as a necessary building block for existing material.
3. Game-based Optimization
To incorporate human expertise, the authors designed an interactive game called "Top-Student," embedded within the XuetangX platform. Students earn points by identifying and "deleting" irrelevant concept suggestions. This crowdsourced "deletion ratio" is fed back into the model as a high-signal feature to refine future results.
Experimental Evaluation
The model was evaluated on four datasets (Data Structure and Algorithm, Introduction to Psychology) across English (Coursera) and Chinese (XuetangX) platforms.
The proposed model, MOOC-CG, significantly outperformed traditional baselines such as PageRank-based SEAL and PU-Learning. On Chinese datasets, the model achieved a Mean Average Precision (MAP) of 0.886, a substantial improvement over the 0.684 achieved by standard graph-based methods. The inclusion of the interactive game component alone provided a +0.14 performance boost, demonstrating that gamified crowdsourcing can effectively augment expert labeling.
Discussion and Insights
A key finding of this research is the power of negative feedback. The "deletion ratio" from the student game proved to be an incredibly strong signal for pruning low-quality suggestions. Furthermore, the use of RNNs to encode the path between concepts in a knowledge graph allowed the model to capture semantic nuances that simple node-link analysis misses.
The researchers noted that Chinese Data Structure courses presented the greatest challenge due to "scattered" concept clusters, which increases the risk of semantic drift. Future work will explore expanding not just related concepts, but specific attributes of those concepts, and testing the framework in less structured domains like Literature or Art History.
Summary updated