We are pleased to share that our latest research paper, “Using Deep Knowledge Tracing to Discover Knowledge Concept Relations and an Explainable Curriculum Structure,” has been accepted for presentation at CSEDU 2026. The camera-ready version has now been submitted, with the work scheduled to be presented at the conference in May.
This milestone reflects an ongoing focus at Adaptemy: advancing adaptive learning not just through performance, but through rigorous, interpretable, and pedagogically meaningful AI.
Why This Research Matters
As digital learning continues to scale, a persistent challenge remains: many systems optimise for prediction, but fail to provide insight. While modern AI models can accurately forecast student performance, they often operate as opaque systems. This creates a disconnect with educators, who require transparency to trust, interpret, and act on these outputs.
Our research addresses a central question:
Can we move from predicting learning outcomes to understanding how learning itself is structured?
This distinction is critical. Without understanding relationships between knowledge concepts, even the most accurate systems risk remaining operationally useful but pedagogically limited.
From Prediction to Structure
The paper builds on Deep Knowledge Tracing (DKT)—a well-established approach that models how student knowledge evolves over time using sequential neural networks. DKT has demonstrated strong predictive performance, but typically does not expose how concepts relate to one another within a curriculum. Our contribution is to extend its utility: Rather than treating the model as an endpoint, we use it as a source of structure.
Specifically, we introduce a method to extract an influence matrix between knowledge concepts—capturing how mastery of one concept affects performance on another. This allows us to:
- Identify latent relationships between skills
- Infer prerequisite structures
- Move toward data-driven curriculum design
Importantly, this approach does not require manually encoding these relationships. Instead, they are learned directly from student interaction data.
A Step Toward Explainable AI in Education
A consistent theme in the literature is the tension between performance and interpretability. Deep learning models capture complex learning patterns, but their “black box” nature limits their usefulness in real educational settings. Educators need more than predictions—they need explanations.
This work positions itself within that gap.
By extracting interpretable structures from a high-performing model, we demonstrate a practical pathway toward explainable AI (xAI) in education—one that aligns with the needs of teachers, researchers, and policymakers.
The framework presented connects:
- AI models (e.g. DKT, LSTM architectures)
- Extraction methods (influence analysis between concepts)
- Educational outcomes (curriculum insight, trust, accountability)
This is not simply a technical refinement; it is a step toward making AI systems evern more powerful in real pedagogical contexts.
Early Findings (and What They Suggest)
Using benchmark educational datasets, we observe that:
- The model-derived curriculum structure differs meaningfully from both
- simple empirical transition models, and
- rigid, human-authored prerequisite graphs
- The extracted structures are denser and more nuanced, capturing relationships that are not explicitly encoded in traditional curriculum maps
- At the same time, there is measurable alignment with established pedagogical structures, suggesting the model is not simply overfitting noise, but identifying meaningful patterns
One interpretation is that DKT acts as a form of signal extraction—filtering observed student behaviour into a probabilistic structure of learning relationships. This opens up a compelling possibility:
Curriculum structure can be discoverable, not just designed.
Implications for Adaptive Learning
For adaptive learning systems, this has direct consequences. Most platforms today rely on:
- static curriculum graphs
- manually defined prerequisite chains
- or heuristic sequencing rules
These approaches are difficult to maintain and often fail to reflect how learners actually progress.
A data-driven, continuously updated curriculum model offers a different path:
- More accurate sequencing of content
- Better identification of learning gaps
- Increased alignment between pedagogy and observed learner behaviour
This aligns closely with the broader evidence base around adaptive learning: systems that personalise effectively can significantly improve learning gain, engagement, and retention.
Looking Ahead to CSEDU 2026
The upcoming presentation at CSEDU 2026 will be an opportunity to engage with the academic and practitioner community on several open questions:
- How should curriculum structures be validated in real-world settings?
- What level of interpretability is sufficient for educators?
- How can these methods be integrated into production learning systems?
We see this work not as a conclusion, but as part of an ongoing effort to bridge learning science, machine learning, and practical deployment
Access to the Research
The paper is not yet publicly available online as it has only recently been submitted. However, we are happy to share it directly with partners and clients on request.
A Broader Perspective
Adaptive learning is often framed as a question of technology. In practice, it is a question of how well systems reflect how people actually learn. This research reinforces a simple but important principle:
The future of AI in education depends not only on better predictions, but on better understanding.
At Adaptemy, this is where we continue to focus our efforts—developing systems that are not only effective, but grounded, interpretable, and aligned with real educational outcomes.
> If your team is committed to improving learning and training with AI, you can book a virtual meeting with Adaptemy here:
