Neural-Symbolic Knowledge Tracing: Injecting Educational Knowledge into Deep Learning for Responsible Learner Modelling

Synthesis: Key limitations exist in both Large Language Models (LLMs)-based tutoring and conventional Deep Knowledge Tracing (DKT):

Combining neural networks with symbolic educational knowledge for interpretable, data-efficient, and temporally reliable learner modeling in intelligent tutoring systems.

The Problem with Pure Neural Approaches

Hooshyar et al. (2026) identify key limitations in both LLM-based tutoring and conventional Deep Knowledge Tracing (DKT):

• Limited adaptivity: LLMs struggle to model learners' evolving knowledge over time — they respond to individual turns rather than tracking skill mastery across sessions.
• Black-box opacity: Deep learning models (LSTM/Transformer-based DKT) lack interpretable learner state representations, making it impossible to audit why a student is flagged as struggling.
• Spurious correlations: Purely data-driven models can latch onto surface patterns (session length, interaction frequency) rather than genuine learning signals.
• Temporal instability: Conventional DKT produces prediction inconsistency rates of 0.43–0.48, meaning predictions frequently "flip-flop" between correct/incorrect as new responses arrive — undermining trust.

The Responsible-DKT Model

Hooshyar et al. propose Responsible-DKT, a neural-symbolic architecture built on the Lifted Relational Neural Networks (LRNN) paradigm using PyNeuraLogic. The key innovation: symbolic educational knowledge is injected as learnable weighted rules that modulate (not override) neural predictions.

Architecture

Injected Educational Knowledge (Symbolic Rules)

1. mastered rule: Two consecutive correct responses on the same skill/quiz → increases prediction confidence. Encodes the pedagogical assumption that sustained correctness signals mastery.
2. not_mastered rule: Three incorrect responses on the same skill/quiz → decreases prediction confidence. Encodes the assumption that repeated failure is a stronger signal — the paper found this rule dominates prediction updates.
3. avg_embed rule: Historical aggregation of past embeddings for the same skill/quiz — provides a long-run average representation of the student's relationship to a knowledge component.

Baselines

• BaseNS-DKT: Identical neural-symbolic architecture but without the educational knowledge rules — isolates the value of symbolic injection.
• PyTorch DKT: Fully data-driven DKT — represents the conventional black-box approach.

Key Findings

Data Efficiency (RQ1)

• >0.80 AUC with only 10% of training data — practically useful for cold-start and sparse-data settings.
• Up to 0.90 AUC with full training data.
• Up to 13% improvement over both baselines.

Temporal Reliability (RQ2)

• Lowest prediction inconsistency rates across all sequence lengths (10, 50, 100, 475):
  • Responsible-DKT: 0.36–0.41
  • BaseNS-DKT: 0.43–0.48, Classic-DKT: 0.44–0.46
• Lower early- and mid-sequence prediction errors.
• Prediction updates remain directionally aligned with observed responses — the model doesn't contradict itself as new data arrives.

Interpretability (RQ3)

• Grounded computation graph exposes the logic behind each prediction — inherently interpretable, not post-hoc explained.
• Supports both local (per-prediction) and global explanations.
• Empirical evaluation of pedagogical assumptions revealed that non-mastery rules (repeated incorrect responses) have a stronger influence on prediction updates than mastery rules — an insight that would be invisible in black-box models.

Dataset

Real-world 6th-grade Maths from Opiq: 167 students, 21,471 interactions, 13 skills, 1,058 quizzes. Scores binarized at first-quartile threshold (37/100).

Connection to Knowledge Tracing Landscape

Traditional Knowledge Tracing (KT) methods:

• IRT-based (Item Response Theory): Interpretable Knowledge Tracing — interpretable difficulty parameters but limited sequential modeling.
• BKT (Bayesian Knowledge Tracing): Explicit prerequisite structures; interpretable but rigid.
• DeepKT (LSTM/Transformer): High predictive capacity but opaque.
• Neural-symbolic KT bridges the gap: retains interpretability of symbolic approaches while leveraging neural networks' representational power.

Implications for Responsible AI in Education

The paper positions this under responsible AI (Goellner et al., 2024: "human-centered approach fostering user trust through ethical decision-making, explainable outcomes, and privacy-preserving implementation"):

1. Transparency: Symbolic components make learner state auditable — critical for safety auditing.
2. Pedagogical validity: Updates follow educational theory (Mastery Learning, repeated-failure signals), not just data patterns.
3. Data efficiency: Practical in real educational settings where large labeled datasets are rare.
4. Temporal trustworthiness: Predictions don't oscillate wildly — a key requirement for student trust and instructor decision-making.

This contrasts with opaque LLM-based approaches in tutoring alignment and demonstrates a path toward A principled way to think about AI in education: guidance for educators and policy makers based on goals, models — augment, don't displace, pedagogical reasoning.

Open Questions

• How does Responsible-DKT scale to larger skill graphs (hundreds of knowledge components)?
• Can the symbolic rule set be extended with domain-specific pedagogical knowledge (e.g., common Misconceptions about AI)?
• How does this compare to LLM-based longitudinal student modeling in both accuracy and interpretability?
• Is the 13% improvement maintained with more diverse student populations and subject domains?

What this means for practice

• Learners. Ask the tutor to show why it flagged you as struggling: Responsible-DKT exposes a computation graph with learned rule weights, so a prediction can be traced to the mastery or non-mastery rule that drove it instead of being taken on trust.
• Developers. Encode pedagogical assumptions explicitly rather than hoping they emerge from data: the reference implementation uses a mastered rule (two consecutive correct responses on a skill), a not_mastered rule (three incorrect responses), and a historical average-embedding rule, each with a learnable weight that modulates rather than overrides the neural prediction.
• Developers. Weight repeated failure more heavily than repeated success in mastery rules — the paper found the non-mastery rule dominates prediction updates, an asymmetry that would be invisible in a black-box model.
• Developers. Design for sparse-data settings: the model reaches over 0.80 AUC with only 10% of training data and up to 0.90 AUC with the full set, so cold-start tutoring can precede a large interaction history.
• Developers. Report temporal stability alongside accuracy: the evaluation measures prediction volatility and inconsistency across sequence lengths, and the rule-augmented model was the most consistent, with inconsistency decreasing further as learning histories grew longer.

Limitations

• The evaluation rests on a single real-world dataset — 6th-grade mathematics from Opiq with 167 students, 21,471 interactions, 13 skills, and 1,058 quizzes — with scores binarized at a first-quartile threshold of 37/100, so generalization across grades, subjects, and populations is untested.
• The symbolic rule set is limited to three simple rules and, by the authors' account, does not reflect the broader capabilities of neural-symbolic approaches or capture richer forms of domain knowledge.
• Comparisons cover a classic DKT baseline and a neural-symbolic variant without rule injection only; attention-based and graph-based knowledge tracing models remain untested against it.
• The injected rules are pedagogically motivated but do not operationalize ethical principles, fairness constraints, or privacy-preserving mechanisms; the authors describe the work as a partial, methodological contribution to responsible AI rather than a comprehensive realization of it.

Citation

Hooshyar, D., Šír, G., Yang, Y., Kärkkäinen, T., Hämäläinen, R., Krivich, E., Cukurova, M., Gašević, D., & Azevedo, R. (2026). Neural-Symbolic Knowledge Tracing: Injecting Educational Knowledge into Deep Learning for Responsible Learner Modelling.