Abstract
The rapid expansion of Massive Open Online Courses (MOOCs) has created a need for automated systems that can map the prerequisite dependencies between educational concepts. This paper introduces PREREQ, a hybrid supervised learning framework that combines generative topic modeling with discriminative neural networks. By leveraging Pairwise Link-LDA to learn latent concept representations and a Siamese Neural Network to determine directionality, PREREQ effectively identifies prerequisite relations from unstructured course content and video sequences. Evaluations on university benchmarks and MOOC datasets demonstrate that PREREQ significantly outperforms existing baselines, offering a scalable solution for automated curriculum planning.
Introduction
In the landscape of modern online education, the volume of available content often overwhelms learners. Navigating this information requires a "prerequisite graph"—a structured roadmap indicating the necessary prior knowledge for any given topic. Traditionally, these graphs were manually curated by experts, a process that is neither scalable nor adaptable to the fast-evolving nature of digital curricula.
Automating the discovery of these relationships is challenging due to the need for "directionality." While many models can identify that two concepts are related, determining which must come first is a non-trivial task. Previous attempts relied heavily on hand-crafted features or struggled to generalize across different academic domains.
Datasets and Data Sources
To train and evaluate the PREREQ framework, the authors utilized two distinct types of educational data:
- University Course Dataset: This benchmark includes 654 courses from major US universities, featuring 861 course-level prerequisite edges and 1,008 manually annotated concept-level pairs.
- NPTEL MOOC Dataset: A novel dataset comprising subtitles from 382 computer science video lectures. A key innovation here is the use of temporal order within video playlists as a proxy for prerequisite relationships, assuming that earlier videos naturally introduce concepts required for later ones.
The PREREQ Framework
The PREREQ model employs a two-stage architecture designed to capture both the semantic content of concepts and their sequential dependencies.
1. Latent Representation via Pairwise Link-LDA
Standard word embeddings often fail to capture the specific pedagogical context of educational materials. PREREQ utilizes Pairwise Link-LDA, a variation of Latent Dirichlet Allocation that incorporates document-level links (e.g., Course A is a prerequisite for Course B). By accounting for these links during the topic modeling phase, the system generates refined latent representations that reflect how concepts are positioned relative to one another in a learning sequence.
2. Directional Learning via Siamese Neural Networks
To address the directionality problem, the latent vectors for a pair of concepts ($C_s, C_t$) are fed into a Siamese Neural Network. This network consists of two identical sub-networks with shared weights. It learns to compute the directional "difference" between the concepts, outputting a probability $P_{c_s, c_t}$ that $C_s$ is a prerequisite for $C_t$.
Key Results and Discussion
PREREQ demonstrates substantial improvements over state-of-the-art baselines like CPR-Recover and MOOC-RF.
- University Benchmark: PREREQ achieved an F-score of 59.68, a significant leap from the 24.54 achieved by CPR-Recover.
- MOOC Dataset: The framework reached an F-score of 60.73, outperforming the MOOC-RF baseline (58.07).
A critical insight from the study is the effectiveness of using video playlists as a training proxy. Despite the inherent noise in chronological sequences, the sheer volume of data available through this method allows the model to learn robust features without extensive manual labeling. Furthermore, the model proved highly data-efficient, maintaining strong performance even when training data was reduced by 60%.
Conclusion
The PREREQ framework provides a powerful, scalable approach to mapping the prerequisite structure of online educational resources. By combining the generative strengths of topic modeling with the discriminative precision of Siamese networks, it overcomes the limitations of previous automated methods. Future research will focus on reducing noise in video-based proxies and expanding the model to incorporate multi-modal data, such as visual cues from lecture slides.
Summary updated