A personalized learning app that adapts to each student’s pace, knowledge gaps, and learning style sounds like an obvious improvement over one-size-fits-all content delivery. It is — when it’s built well. The gap between a learning app that calls itself “personalised” and one that delivers genuine adaptive learning is substantial, and most learners can feel the difference after a few sessions. This guide covers what genuine personalisation in learning apps looks like technically, the AI and algorithm approaches that drive it, and how to build the infrastructure to deliver it at scale.
What Personalised Learning Actually Means
Personalisation in learning apps operates at several levels, and understanding which level you are implementing is essential for setting realistic expectations:
- Content sequencing: Choosing which topic or concept to present next based on what the learner has mastered and what they are ready to learn next
- Difficulty adaptation: Adjusting the difficulty of questions or exercises within a concept based on current performance
- Format personalisation: Presenting the same concept in different formats (video, text, interactive exercise) based on which formats a learner engages with and retains best
- Pacing control: Allowing learners to move through content at their own pace, with the system supporting both faster and slower progress appropriately
- Remediation: Detecting knowledge gaps and providing targeted practice on weak areas before advancing to content that depends on those foundations
Many apps that claim personalisation implement only pacing control — the learner moves at their own speed through fixed content. This is better than a fixed pace, but it is not adaptive personalisation. True adaptive personalisation changes what content is delivered, not just when.
The Personalised Learning App Algorithm: Knowledge Tracing
Bayesian Knowledge Tracing
Bayesian Knowledge Tracing (BKT) is the most established algorithm for modelling student knowledge in personalized learning apps. It models each knowledge component (a specific skill or concept) as a binary latent variable — the student either knows it or doesn’t — and updates the probability that the student has mastered the component based on their responses to questions. Four parameters define a BKT model: prior probability of knowing the concept before any practice, probability of learning the concept in any given practice opportunity, probability of making a careless error despite knowing, and probability of guessing correctly despite not knowing.
BKT is computationally efficient (it can run in real time on every learner response) and interpretable (you can explain why content was selected). Its limitation is that it models each knowledge component independently — it doesn’t capture the dependencies between concepts or the fact that a student’s performance on one concept predicts their performance on related concepts.
Deep Knowledge Tracing
Deep Knowledge Tracing (DKT) uses LSTM neural networks to model student knowledge across all knowledge components simultaneously. Unlike BKT, DKT captures dependencies between concepts — a student who struggles with fraction division is likely to struggle with ratio problems, and DKT will predict this even before the student has encountered ratio problems. DKT consistently outperforms BKT on prediction accuracy but requires more training data and is less interpretable (you cannot easily explain why DKT recommended a specific piece of content).
For a new personalized learning app with limited historical data, BKT is the practical starting point. As you accumulate interaction data across thousands of learners, DKT becomes viable and the prediction quality improvement justifies the additional complexity.
Content Recommendation: Translating Knowledge State to Content Selection
Knowledge Graph and Prerequisite Mapping
A knowledge graph defines the dependencies between concepts in your learning domain: concept A must be mastered before concept B can be learned; concepts C and D can be learned in either order; concept E integrates concepts A, C, and D. This graph is what allows the personalized learning app to sequence content intelligently — the system will not present a concept whose prerequisites haven’t been mastered, and it will identify the most productive next concept based on the learner’s current knowledge state and the graph structure.
Building this knowledge graph is intellectual work that domain experts must do — it cannot be automated. AI can assist in identifying likely dependencies from learning outcome data, but the graph structure must be validated by subject matter experts who understand the domain’s conceptual dependencies.
The Recommendation Policy
Given a knowledge graph and a knowledge state estimate for each learner, the recommendation policy decides what to present next. Common policies: always present the highest-probability-of-learning-gain concept (greedy exploitation), balance between concepts the learner is almost-but-not-quite mastered (high reinforcement value) and new concepts the learner is ready to encounter (exploration), and prioritise concepts that are blocking progress on downstream content the learner is working toward (prerequisite filling).
Multi-armed bandit approaches — specifically contextual bandits — provide a principled way to balance exploration (trying new content types) and exploitation (using what works) while adapting to individual learner preferences. Thompson Sampling is the most commonly used bandit algorithm in learning apps due to its natural handling of uncertainty.
Spaced Repetition: The Most Proven Personalisation Technique
Spaced repetition is the most well-validated personalisation technique in educational research: reviewing content at increasing time intervals based on memory strength produces dramatically better long-term retention than massed practice. The SuperMemo SM-2 algorithm (the basis for Anki and Duolingo’s early algorithm) calculates the optimal review interval for each concept based on the learner’s historical performance on that concept. Implementing spaced repetition correctly requires: tracking the last review time and performance for every concept-learner pair, scheduling future review sessions appropriately, and presenting review content alongside new learning content rather than separating them into distinct “review sessions” that learners skip.
AI-Powered Features Beyond Knowledge Tracing
Natural Language Feedback
LLMs (Claude, GPT-4o) can provide detailed, personalised feedback on open-ended responses — explaining why an answer was partially correct, identifying the specific misconception behind a wrong answer, and suggesting the correct reasoning process. This type of feedback was previously only possible from human tutors. For learning apps covering subjects with clear correct/incorrect reasoning (mathematics, programming, science), AI-generated explanatory feedback provides significant learning value at a cost that scales with the number of learners rather than requiring proportional human tutor time.
Implementation: route open-ended responses to an LLM with context about the learning objective, the question, and the correct answer. Parse the LLM output for the feedback components you want to display. Cache feedback for common wrong answers to reduce API cost and latency.
Difficulty Calibration
Item Response Theory (IRT) models the difficulty of each question and the ability of each learner on the same scale, allowing the system to select questions at the right difficulty level — challenging enough to promote learning, easy enough to maintain motivation. IRT requires calibration data (student responses to questions) to estimate item difficulty parameters. For new apps, approximate initial difficulty estimates from content experts, and refine them as response data accumulates.
Data Infrastructure for Personalised Learning Apps
Personalisation quality is bounded by data quality. Every learner interaction — question attempted, time spent, hint used, response submitted — should be captured as an event in a structured learner data store. The xAPI (Experience API) standard defines a common format for learning event data (“Actor did Verb Object”) that enables interoperability with analytics tools and future algorithm improvements. Build your data capture around xAPI events from the beginning; retrofitting structured data capture onto a system that wasn’t designed for it is consistently painful.
For related content on building learning apps with offline-first architecture — essential for mobile learners in variable connectivity environments — see our post on Quick Kids Learning Apps: Offline-First Design Best Practices.
Frequently Asked Questions
How much data do I need before personalisation improves on a fixed curriculum?
BKT-based personalisation shows improvement with as few as 10–20 learner interactions per knowledge component. At fewer interactions than this, the knowledge state estimate is uncertain enough that personalised and fixed sequencing produce similar results. As data accumulates across learners, the model improves for new learners (cold start) through population-level parameter estimation. Expect 500–1,000 learners completing substantial content before population-level model quality is meaningfully better than a well-designed fixed curriculum.
How do you handle the cold start problem for new learners?
Before any interaction data exists for a new learner, use population-level defaults (the average knowledge state for learners at a similar stage) as the starting estimate. Optionally, use a short diagnostic assessment at onboarding — 10–15 carefully selected questions covering the key concepts — to rapidly calibrate the initial knowledge state estimate before the learner begins the main content sequence.
How do you measure whether personalisation is actually working?
The gold standard is a randomised controlled trial: assign learners randomly to personalised and non-personalised conditions and compare learning outcomes. In practice, most apps use a combination of engagement metrics (session length, return rate, completion rate) as leading indicators and learning assessment scores as the ultimate outcome metric. Be careful not to optimise purely for engagement — a learning app can be highly engaging without producing strong learning outcomes.
Conclusion
Building a genuinely personalised learning app requires investment in three areas: the algorithmic infrastructure (knowledge tracing, recommendation policy, spaced repetition), the content infrastructure (knowledge graph, well-tagged content, difficulty-calibrated questions), and the data infrastructure (structured event capture for every learner interaction). Apps that invest in all three deliver measurably better learning outcomes than fixed curricula. Apps that invest in only one or two usually produce personalisation theatre — the appearance of adaptation without the substance.
If you are building a personalized learning app and want to discuss the algorithmic approach or technical architecture for your specific subject domain, get in touch with the Lycore team.
