💻 computer science

Tracing intermittent knowledge states based on knowledge-context-aware augmentation of Q-matrices

This paper proposes a knowledge tracing model that utilizes knowledge-context-aware Q-matrix augmentation and topological sorting of knowledge concepts to effectively trace students' intermittent knowledge states, which exhibit short-term stability and long-term variability, validated through experiments on multiple datasets.

Original authors: Lankun Guo, Jianghui Yi, Jinhai Li, Zhiqiang Yuan

Published 2026-07-31

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Lankun Guo, Jianghui Yi, Jinhai Li, Zhiqiang Yuan

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to track the progress of a student learning a complex subject, like mathematics. In the world of "knowledge tracing," scientists act like detectives, trying to figure out exactly what a student knows and doesn't know at any given moment. They do this by looking at how students answer questions over time. Traditionally, many computer models assumed that a student's knowledge is like a shaky tightrope walker: constantly wobbling, changing, and fluctuating wildly with every single question they answer. If a student gets one question right, their "knowledge level" jumps up; if they get one wrong, it crashes down.

However, real life isn't quite that chaotic. Think of learning like building a house. You might spend a week laying bricks (learning), and then take a weekend off (forgetting a little bit), but when you come back on Monday, the foundation you built last week is still mostly there. It doesn't vanish just because you took a break. Most existing computer models miss this "intermittent" nature of learning—the idea that knowledge stays stable for a while during a test, but can shift significantly between different testing periods. This paper, titled "Tracing intermittent knowledge states based on knowledge-context-aware augmentation of Q-matrices," aims to build a smarter detective that understands this rhythm. It introduces a new model called INKT that doesn't just watch the student; it understands the map of the knowledge itself and how students learn and forget over time.

The Map and the Missing Pieces

To understand the paper's solution, we first need to look at the tools the researchers used. Imagine a giant map of a city where every street represents a "knowledge concept" (like algebra or geometry) and every building is a "test question." In this city, some streets lead directly to others; you can't drive to the highway without first taking the on-ramp. This map is called a Q-matrix. It's a list that tells the computer which concepts are needed to answer which questions.

The problem is that the maps provided by human experts are often incomplete. They only show the obvious connections. For example, a map might say, "To solve this calculus problem, you need to know derivatives." But it forgets to mention that to know derivatives, you also need to know limits, and to know limits, you need to know functions. The computer, looking at the incomplete map, thinks the student only needs to know derivatives. If the student gets the calculus problem wrong, the computer blames the calculus skill, missing the fact that the student actually forgot how to handle limits.

The authors of this paper realized that to fix the map, they needed to look at the "topology" of the city—the shape of the connections. They proposed a way to "augment" (or upgrade) the Q-matrix. Instead of just looking at the direct links, their algorithm traces the complete paths through the knowledge graph. If Concept A leads to Concept B, and Concept B leads to Concept C, the computer learns that doing a task for C also involves A and B. They used a mathematical method called "topological sorting" to organize these concepts in a logical order, like arranging books on a shelf from the easiest to the hardest. By doing this, they could fill in the "hidden" connections, turning a sparse, incomplete map into a rich, detailed guide that shows how all the concepts are secretly related.

The "Intermittent" Detective

Once they had a better map, the researchers built their new detective model, INKT. The core idea here is that knowledge states are "intermittent." This means that during a short test (like a 30-minute quiz), a student's knowledge is relatively stable. They aren't suddenly forgetting everything or learning a new chapter in the middle of the quiz. However, between tests—say, a week apart—students might forget some things (the "forgetting rate") or learn new things (the "learning rate").

Most old models treated every question as a brand-new event where the student's knowledge could swing wildly. INKT, on the other hand, respects the time windows. It assumes that if you take a test today, your knowledge is a slightly evolved version of what it was yesterday, adjusted by how much you learned or forgot in the meantime. It uses a mathematical framework called Item Response Theory (IRT), which is a fancy way of saying it considers not just if you got the answer right, but how hard the question was and how good the question is at telling the difference between a smart student and a guessing student.

The model also accounts for the "forgetting curve," a concept from psychology that suggests we lose information over time if we don't use it. INKT calculates a specific "forgetting rate" for each student and each concept, and a "learning rate" for how fast they pick things up. It then uses these rates to predict what the student's knowledge state should be at the next test, rather than just guessing based on the last answer.

What the Experiments Showed

To see if their new detective was any good, the researchers tested it on several datasets. They created a custom dataset called MTest from a real high school in Shenzhen, China, involving 69 students, 7 tests, and 88 knowledge concepts. They also used three public datasets (Math1, Math2, and FrcSub) and even generated a massive simulated dataset called FMTest using a method called Formal Concept Analysis to see how the model handled large groups of students.

The results were quite promising. When they compared INKT to other models (some of which were simpler, and some of which were more complex but didn't account for the "intermittent" nature of learning), INKT consistently performed better.

On the MTest dataset, INKT achieved an accuracy (ACC) of around 0.77 in the best test scenario, while the next best model only reached about 0.57.
On the large simulated FMTest dataset, INKT hit an accuracy of 0.94, significantly outperforming the others which hovered around 0.81.
The model also showed it was very stable. Even when they changed the number of students or the mix of tests, INKT didn't crash; it kept giving accurate predictions.

The researchers also ran "ablation" studies, which are like taking the engine apart to see which part does what. They found that if they removed the "forgetting rate" or the "learning rate" from the model, the performance dropped significantly. This proved that tracking how students learn and forget over time is essential. They also found that using the "augmented" Q-matrix (the upgraded map with hidden connections) was crucial. When they used the old, incomplete map, the model wasn't as good at predicting who would get what right.

The Verdict

The paper concludes that by combining a smarter map of knowledge (the augmented Q-matrix) with a model that respects the rhythm of learning and forgetting (the intermittent state tracing), we can get a much clearer picture of what students actually know. The authors suggest that this approach is more rational than models that assume knowledge is constantly and wildly fluctuating.

However, the authors are careful not to claim this is a perfect, solved problem. They admit that their "map" of knowledge connections was still built by human experts, which can be subjective. They also note that their model assumes a student's current knowledge only depends on the immediately previous state, ignoring more complex "attention" mechanisms that might be at play. They suggest that future work could involve automatically generating these knowledge maps from data and using more advanced attention mechanisms to make the model even more accurate.

In short, this paper offers a playful but rigorous way to stop treating students like jittery tightrope walkers and start treating them like builders who lay bricks, take a break, and come back to find their foundation mostly intact. By understanding the hidden connections in the curriculum and the natural ebb and flow of memory, the INKT model provides a more honest and accurate view of the learning journey.

Technical Summary: Tracing Intermittent Knowledge States Based on Knowledge-Context-Aware Augmentation of Q-Matrices

Problem Formulation
Existing knowledge tracing (KT) models often assume that a student's knowledge state fluctuates continuously and significantly with every exercise interaction. However, in practical educational scenarios, students typically engage in learning in batches (e.g., during specific exams or test windows) separated by longer intervals. During these intervals, latent learning and forgetting occur, while knowledge states remain relatively stable within short time windows (e.g., during a single exam). Current deep learning-based models often fail to capture this "intermittent" nature of knowledge state evolution, treating time series data as a continuous stream of constant change. Furthermore, traditional cognitive diagnosis models rely heavily on Q-matrices (mapping exercises to knowledge concepts) that are often provided by experts and contain only explicit connections, missing implicit relationships. This incompleteness reduces tracing accuracy, while the "black box" nature of deep learning models often sacrifices interpretability.

Methodology
The authors propose INKT (Intermittent Knowledge Tracing), a model designed to trace students' knowledge states by admitting characteristics of stability in short time ranges and variability across different stages. The methodology consists of three core modules:

Topological Index Mining:
The model utilizes a knowledge concept graph $G$ where nodes represent concepts and directed edges represent preparatory relationships (prerequisites). The authors define a set of "complete paths" within this graph. They introduce a topological index ( $x_k$ ) for each knowledge concept to represent its relative position in a desired linear order. This index is optimized by maximizing a function that balances:
- Positive Correlation: The alignment between the topological index and the accumulated preparatory degree (derived from edge weights) along complete paths.
- Penalty Mechanism: A term that penalizes contradictions where a prerequisite concept has a lower index than the concept it supports.
  This optimization is solved via gradient descent to derive a linear ordering of concepts that best fits the graph structure.
Knowledge-Context-Aware Q-Matrix Augmentation:
Using the derived topological indices and the knowledge concept graph, the model augments the original sparse Q-matrices. Instead of relying solely on explicit expert labels, the model infers implicit knowledge concepts for each exercise.
- For an exercise $e$ , the set of related concepts includes both explicit ones and those appearing earlier in the complete paths of the explicit concepts.
- The correlation values ( $\tilde{q}_{ek}$ ) are quantified by optimizing a Bayesian objective function. This function maximizes the compatibility between the augmented Q-matrix values and the topological indices of the knowledge concepts, ensuring the updated matrix reflects the underlying cognitive structure.
Intermittent Knowledge Tracing Model:
The core tracing module integrates Item Response Theory (IRT) with a recursive state update mechanism based on Bayesian formulas.
- State Dynamics: The mastery of a student $s$ on concept $k$ at stage $i$ ( $I^{(i)}_{sk}$ ) is modeled as a function of the previous stage's mastery, weighted by a learning rate ( $c^{(i)}_{sk}$ ) and a forgetting rate ( $f^{(i)}_{sk}$ ). The forgetting rate is modeled using the Ebbinghaus forgetting curve, dependent on the time interval $\Delta t$ .
- Proficiency and Scoring: A student's proficiency on an exercise is calculated as the weighted sum of their mastery on related concepts (using the augmented Q-matrix). The simulated scoring rate is then derived using a 2-parameter IRT logistic function involving exercise difficulty and discrimination.
- Optimization: The model learns parameters (initial mastery, learning/forgetting rates, item difficulty/discrimination) by minimizing a loss function derived from the log-posterior distribution, utilizing gradient descent (specifically Adam optimization).

Key Contributions

Intermittent State Modeling: The paper introduces a framework specifically designed for scenarios where knowledge states are stable within short testing windows but evolve across longer intervals, addressing a gap in existing continuous-time KT models.
Q-Matrix Augmentation via Topology: A novel method is proposed to automatically augment sparse Q-matrices by leveraging the topological structure of the knowledge concept graph. This captures implicit knowledge dependencies without requiring additional expert labeling.
Interpretability and Rationality: Unlike many deep learning "black box" models, INKT explicitly models cognitive parameters (learning/forgetting rates) and item attributes (difficulty/discrimination), providing a transparent and theoretically grounded explanation of student performance.
Recursive Update Mechanism: The model employs a recursive formula that dynamically updates knowledge states based on time intervals and personalized learning/forgetting rates, moving beyond the static assumptions of classical IRT.

Experimental Results
The authors evaluated INKT on three public datasets (FrcSub, Math1, Math2) and two newly collected datasets (MTest and a formal concept analysis-generated FMTest).

Performance: INKT consistently outperformed baseline models (including 1-parameter and 3-parameter IRT variants, and fuzzy-set based models) across metrics such as Accuracy (ACC), Mean Squared Error (MSE), and Mean Absolute Error (MAE).
Robustness: The model demonstrated superior stability under varying data splits (different proportions of tests and students) and showed strong generalization on large-scale datasets.
Ablation Studies: Removing either the learning rate or forgetting rate components significantly degraded performance, confirming the necessity of modeling both cognitive factors. Replacing the augmented Q-matrix with the original one also resulted in performance drops, validating the value of the augmentation strategy.
Case Studies: Visualizations of specific students' mastery trajectories revealed the model's ability to capture nuanced learning patterns, such as persistent weaknesses in specific concepts or gradual improvements, and the dynamic interplay between learning and forgetting.

Significance and Claims
The paper claims that INKT provides a more rational and accurate approach to knowledge tracing by aligning with the actual cognitive processes of students (intermittent stability and variability). By integrating topological analysis for Q-matrix augmentation and explicit cognitive parameters for state tracing, the model overcomes the static limitations of classical IRT and the interpretability issues of deep learning models. The authors assert that their approach offers a "three-dimensional" (student-time-item) dynamic modeling framework that effectively simulates the temporal evolution of knowledge states, making it particularly suitable for personalized education systems where understanding the process of learning is as important as predicting the outcome.

Limitations and Future Work
The authors acknowledge that the preliminary relationships in the knowledge graph are currently provided by experts, introducing potential subjective bias. Future work aims to develop methods for automatically extracting knowledge graphs from data. Additionally, the current recursive mechanism assumes the current state depends only on the previous state; the authors suggest incorporating attention mechanisms to capture more complex dependencies as a future direction. Finally, they note that the current Q-matrix update strategy focuses on linear ordering, whereas more complex logical structures between concepts may require further investigation.

The Map and the Missing Pieces

The "Intermittent" Detective

What the Experiments Showed

The Verdict

Technical Summary: Tracing Intermittent Knowledge States Based on Knowledge-Context-Aware Augmentation of Q-Matrices

More like this