8.4.1 Categories of Artificial Intelligence Techniques

The AI techniques applied in learning analytics span multiple paradigms, each originating in different theoretical traditions and suited to different kinds of educational problems. A learning engineer who understands this taxonomy can make principled choices about which technique to apply to a given problem, rather than defaulting to the most recent or most hyped approach.

Rule-Based AI: Production Systems and ACT-R

The oldest AI paradigm still in active use in learning analytics is rule-based AI — systems that encode expert knowledge as explicit conditional rules (if-then productions) and apply them to input data. The theoretical foundation for this approach in education is the ACT-R cognitive architecture, developed by John R. Anderson at Carnegie Mellon.[LE-LS-PP-002]

ACT-R models cognition as a set of production rules that fire when their conditions are satisfied, producing changes in working memory and triggering actions. In intelligent tutoring systems, the "expert model" — the model of correct domain knowledge that the ITS uses to diagnose learner errors — is typically expressed as a production-rule set. Koedinger and Anderson's foundational work demonstrated that cognitive task analysis (the systematic decomposition of a skill into its constituent production rules) provides the knowledge-component structure that drives both adaptive problem selection and error-specific feedback.[LE-LS-AP-001]

The Cognitive Tutors deployed in thousands of schools were built on this architecture: a production-rule expert model generated solution paths, and the student model (BKT) tracked mastery of each production rule independently.[LE-LS-AP-014] Rule-based AI systems of this kind are highly interpretable — every decision can be traced to a specific rule — but they require extensive knowledge engineering effort to build and are difficult to scale to domains where expert knowledge is tacit or not well-formalized.

SOAR, the unified cognitive architecture developed by Laird, Newell, and Rosenbloom, extends the production-rule paradigm with a chunking mechanism that models procedural learning itself: new production rules are created by chunking together the problem-solving traces of successfully solved problems.[LE-LS-AP-015] This makes SOAR relevant not just as an expert model but as a computational theory of how skills are acquired through practice.

Bayesian AI: Probabilistic Student Modeling

Bayesian AI techniques represent knowledge and uncertainty as probability distributions and update beliefs as new evidence arrives. In learning analytics, this paradigm is instantiated most directly in Bayesian Knowledge Tracing, which uses Bayes' theorem to update the probability of skill mastery after each practice response.[LE-LS-AP-011]

Beyond BKT, the Bayesian paradigm encompasses Dynamic Bayesian Networks (which model temporal dependencies between knowledge states), Bayesian networks for misconception diagnosis (which represent the causal structure of prerequisite knowledge), and Bayesian models of learner affect. These approaches share the virtue of interpretability — the probability distributions are meaningful to domain experts and can be communicated to learners and teachers — and the limitation of requiring explicit specification of the model structure, which may be difficult when the knowledge domain is complex or poorly understood.

Deep Learning: Neural Knowledge Tracing and Beyond

Deep learning approaches replace the explicit model structures of rule-based and Bayesian AI with neural network architectures that learn representations directly from data. The most influential application in learning analytics is Deep Knowledge Tracing (DKT), which uses recurrent neural networks to model the sequence of learner responses and predict future performance. Published research and presentations at the IEEE Transactions on Learning Technologies have documented that DKT can outperform BKT on prediction accuracy, particularly when training data is large.[LE-LS-JO-003]

The trade-off is interpretability: deep learning models are often difficult to explain, making it hard to translate their outputs into specific instructional decisions or to audit them for fairness. The AIED Conference has been the primary venue for advances in interpretable deep learning for education, including attention mechanisms that highlight which prior interactions most influenced a prediction.[LE-LS-CE-004]

Reinforcement Learning for Adaptive Sequencing

Reinforcement learning (RL) treats the tutoring or adaptive system as an agent that must learn a policy for selecting instructional actions (problems, hints, content items) in order to maximize a learner outcome signal (mastery, engagement, test performance). Unlike supervised learning, RL does not require labeled examples of correct decisions — it learns from the feedback signal generated by the learner's responses to the system's choices.

RL-based adaptive sequencing has been demonstrated in experimental deployments, where it was shown to improve learning efficiency relative to fixed curricula or simpler adaptive rules. The Learning Engineering Toolkit notes that RL approaches require careful design of the reward signal — what the system is optimizing for — because misspecified rewards can produce systems that optimize for superficial engagement or test-taking strategies rather than durable learning.[LE-LS-GL-007]

Natural Language Processing for Discourse Analysis

NLP techniques enable learning analytics to process the text of learner responses, discussion forum posts, and open-ended assessments — a vast and largely untapped data source in most learning systems. Applications include automated scoring of short-answer and essay responses, discourse analysis of collaborative learning interactions, sentiment analysis of forum discussions, and error analysis of written explanations.

Generative AI — large language models capable of generating contextually appropriate text — extends NLP from analysis to production: these systems can generate feedback, produce explanations at different levels of detail, and engage in Socratic dialogue with learners. The AIED Conference has become the primary venue for evaluating generative AI in educational contexts, with particular attention to accuracy, hallucination risks, and equity implications.[LE-LS-CE-004]

Community Recommended Reading

Chapter 6: LE Toolkit — Learning Engineering Applies the Learning Sciences — Jim Goodell, Janet Kolodner, Aaron Kessler (2022)

Source: goodell-ch6-le-applies-learning-sciences · confidence: medium · expert-validated: false

Context enrichment applied 2026-04-17. See wrgr/lecommons for source data.