Data-Driven Discovery of Educational Prerequisite Graphs

This direction is merged into another direction: prerequisite-aware recommendation. Its last version stays readable and citable here.

Version 1 of 126 papers30 conceptsfrozen 2026-10-04

In traditional education, the "map" of a subject—which concepts must be mastered before others—is usually drawn by human experts. These experts create a "Q-matrix," a manual mapping of exercises to skills. However, manual mapping is expensive, slow to update, and often misses the subtle, informal dependencies that actually dictate how students learn.

Recent research has shifted toward automated prerequisite discovery: using student performance data, textbook structures, and large language models to mathematically infer the hidden "prerequisite graph" of a domain.

The Directional Nature of Learning

The core intuition behind automated discovery is that learning is fundamentally directional. If Concept A is a prerequisite for Concept B, then a student's mastery of A should be a strong predictor of their future success in B, but the reverse is rarely true.

Andrew Lan, Christoph Studer, and Richard Baraniuk formalized this in Time-varying Learning and Content Analytics via Sparse Factor Analysis by applying a lower-triangular constraint to state transition matrices. By forcing the model to assume that knowledge flows in one direction, the system can automatically recover prerequisite structures from longitudinal performance data. This structural assumption mirrors the pedagogical reality: you cannot master calculus without first mastering algebra.

Inferring Causality from Performance

If we treat skills as "causes" of success, we can use causal discovery algorithms to map hierarchies. Shenbao Yu and Yifeng Zeng proposed the CSPS framework, which treats prerequisite skills as exogenous causes and dependent skills as endogenous effects. By analyzing student success across different exercises, the model identifies which skills act as the "gatekeepers" for others (Causal-Driven Skill Prerequisite Structure Discovery).

This signal is often hidden in the "item order effects" of randomized problem sets. Zachary A. Pardos and Neil T. Heffernan demonstrated that by varying the order of questions, researchers can detect when the mastery of one item significantly boosts the learning rate of another, revealing dependencies that even instructional designers might overlook (Determining the Significance of Item Order In Randomized Problem Sets).

Mining Content and Language Models

Beyond student performance, the way we organize information in text provides a secondary signal for prerequisite discovery. Shivam Pal and Pawan Goyal showed that the structured flow of textbooks—combined with the transitive property (if A leads to B and B leads to C, then A is a prerequisite for C)—can bridge gaps where foundational concepts are assumed but not explicitly linked (Finding Prerequisite Relations between Concepts using Textbook).

More recently, Large Language Models (LLMs) have proven capable of zero-shot concept graph recovery. Rui Yang and colleagues found that LLMs can predict prerequisite relationships between concept pairs by leveraging their internal world knowledge, effectively reconstructing domain-specific hierarchies without needing traditional graph embeddings (Leveraging Large Language Models for Concept Graph Recovery and Question Answering in NLP Education).

Latent Structures in Deep Learning

Deep learning models for knowledge tracing, such as Deep Knowledge Tracing (DKT), can also "hallucinate" these graphs. Chris Piech and Jonathan Huang showed that by analyzing how a model's prediction for one skill changes after a student masters a different skill, the model autonomously maps the underlying curriculum structure (Deep Knowledge Tracing).

Similarly, the Knowledge Query Network (KQN) developed by Jinseok Lee maps skills into a shared vector space. The probabilistic similarity between these vectors allows researchers to infer prerequisite relationships: skills that are "close" in this latent space and share a temporal dependency are likely linked in a prerequisite chain (Knowledge Query Network for Knowledge Tracing).

Why It Works: The Collective Signal

The power of these methods lies in the collective behavior of learners. As Aixin Sun and Jie Zhang observed in massive open online courses (MOOCs), the transition sequences of millions of students reveal a "heartbeat" of learning. When thousands of students consistently take Course A before Course B to succeed, they are providing a high-confidence signal of a latent dependency that no manual catalog could capture as accurately (The Crowd in MOOCs: A Study of Learning Patterns at Scale).

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_data_driven_discovery_of_educational_pre_74b51e55,
  title = {Data-Driven Discovery of Educational Prerequisite Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-driven-discovery-of-educational-prerequisite-graphs/rdr_74b51e55ccea41268af1b1e988295266},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_data_driven_discovery_of_educational_pre_74b51e55_v1,
  title = {Data-Driven Discovery of Educational Prerequisite Graphs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/data-driven-discovery-of-educational-prerequisite-graphs/rdr_74b51e55ccea41268af1b1e988295266/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}