PreBiGE: Course Recommendation Using Course Prerequisite Relation Embedding and Bipartite Graph Embedding

Hafsa Kabir Ahmad, Bo Liu, Bello Ahmad Muhammad, Mubarak UmarJournal of Networking and Network Applications 2022 · DOI

TL;DR

PreBiGE is a novel recommendation framework for Massive Open Online Courses (MOOCs) that combines traditional collaborative filtering with course prerequisite dependencies. By jointly embedding student-course interactions and course-to-course prerequisite relations extracted from a knowledge graph, the model significantly improves recommendation accuracy and relevance.

What Problem Does This Paper Address?

In the wake of the COVID-19 pandemic, online learning has exploded, leaving students overwhelmed by thousands of available courses. Traditional recommender systems usually rely on Collaborative Filtering (CF)—the "people who liked X also liked Y" approach.

However, education is unique: you can't just take "Advanced Quantum Mechanics" because it's popular; you need to take "Linear Algebra" first. Most existing graph-based recommenders capture high-order "collaborative signals" (e.g., students who take similar sets of courses) but ignore the prerequisite dependencies that dictate a logical learning path. The authors address this gap by integrating these structural dependencies into the recommendation engine.

Methods

The authors propose PreBiGE (Prerequisite and Bipartite Graph Embedding), which operates in three main stages:

  1. Prerequisite Extraction via Knowledge Graphs: The model builds a heterogeneous knowledge graph containing students, courses, and concepts. It uses meta-paths (sequences of relations like Course A $\to$ Concept 1 $\to$ Concept 2 $\to$ Course B) to infer if one course is a prerequisite for another. The number of paths between courses determines the "strength" (weight) of the prerequisite link.
  2. Graph Construction:
    • A Bipartite Graph captures direct student-course interactions.
  3. A Course Prerequisite Graph captures the logical flow between subjects.
  4. Homogeneous Graphs are projected to capture high-order collaborative signals (student-student and course-course similarities).
  5. Joint Embedding & Optimization: Using random walks and a Skip-gram model (similar to DeepWalk), PreBiGE learns low-dimensional representations (embeddings) for students and courses. The optimization function jointly balances three factors:
  6. Direct student-course interactions.
  7. High-order collaborative relationships.
  8. Course prerequisite dependencies.

Datasets

The model was evaluated using the MOOCCube dataset, a large-scale repository from the XuetangX MOOC platform.

  • Stats: 884 students, 307 courses, and 17,614 knowledge concepts.
  • Relations: Includes ~5,900 student-course interactions and ~900 concept-to-concept prerequisite relations.

Key Insights and Ideas

  • Prerequisites as a Regularizer: The "aha moment" is that prerequisite relations act as a constraint on the embedding space. They ensure that courses which are logically linked are positioned closer together in the vector space, even if they don't have many shared students yet.
  • Weighted Dependencies Matter: The authors found that treating all prerequisites as equal is less effective than weighting them. A higher number of concept-level meta-paths between two courses indicates a "stronger" prerequisite relationship, which leads to better recommendations when modeled.
  • The Sparsity Solution: By using high-order collaborative signals (looking at indirect connections in the graph), the model overcomes the "cold start" or sparse data problem where many students have only taken one or two courses.

Key Results

PreBiGE was compared against several state-of-the-art baselines, including BPRMF, NeuMF, and graph-based models like NGCF and LightGCN.

  • Superior Performance: PreBiGE outperformed the best baseline (LightGCN) across all metrics.
  • Quantitative Gains: It achieved a 3.61% improvement in F1@10 and a 1.38% improvement in MRR@10 (Mean Reciprocal Rank) over the strongest baseline.
  • Ablation Study: Removing the prerequisite component (the $PreBiGE_{WOP}$ variant) led to a significant drop in performance, proving that prerequisite knowledge is a critical feature for educational recommendations.

Limitations and Future Work

While PreBiGE successfully incorporates prerequisite logic, the authors note a few areas for growth:

  • Temporal Dynamics: The current model doesn't fully account for the chronological order in which a specific student interacts with courses over time.
  • User Behavior: Future iterations could benefit from deeper modeling of student behavior (e.g., video pause/rewind patterns) to better understand their mastery of specific concepts before recommending the next course.
  • Scalability: As knowledge graphs grow to millions of concepts, the computational cost of mining meta-paths may become a bottleneck.

Summary updated

Abstract

A growing number of students enrol in online education to improve their skills. However, students are faced with the challenge of finding courses that meet their individual needs. Recommender systems were introduced to help students choose the courses that best meet their needs. To learn better representations of students and courses for improved recommendation results, existing graph-based recommender systems utilize the high-order collaborative signals between set of students or set of courses from a bipartite graph. However, courses also have prerequisite dependency between them, which when utilized together with collaborative relations can improve recommendation results. On this basis, we propose a model that utilizes the high-order relation between set of courses, the prerequisite dependency between courses, as well as the direct relation between students and courses. Using meta-paths generated from the knowledge graph, our model extracts the prerequisite dependency between courses, which is then used to generate a course prerequisite graph. The course prerequisite graph and the student-course bipartite graph are used to learn the representation of the students and courses, jointly capturing the prerequisite dependency, high-order collaborative relations as well as direct relations. The learned representations are used for recommendation. The experiments on real-world dataset show the superiority of our proposed method, achieving 3.61% on F1@10 and 1.38% on Mrr@10.
BibTeX
@article{ahmad2022prebige,
  author = {Hafsa Kabir Ahmad and Bo Liu and Bello Ahmad Muhammad and Mubarak Umar},
  title = {{PreBiGE: Course Recommendation Using Course Prerequisite Relation Embedding and Bipartite Graph Embedding}},
  journal = {Journal of Networking and Network Applications},
  year = {2022},
  doi = {10.33969/j-nana.2022.020404},
}
Figure