Lesson 2: The Learner Model in Adaptive Learning Systems

2.1 Introduction

In adaptive learning systems, the Learner Model serves as a dynamic representation of an individual learner’s current state of knowledge, skills, abilities, and other relevant attributes. It is the core component that enables personalization by informing the system about the learner’s progress and guiding the selection of appropriate instructional strategies.

This lesson delves deeply into the mathematical and computational techniques used to construct and update the learner model. We will explore various modeling approaches, including probabilistic models like Bayesian Knowledge Tracing and psychometric models like Item Response Theory. Additionally, we will discuss the challenges and considerations in designing an effective learner model.

2.2 Objectives

By the end of this lesson, you should be able to:

• Understand the purpose and importance of the learner model in adaptive learning systems.

• Describe different methods of representing a learner’s knowledge and skills.

• Apply mathematical models to estimate and update a learner’s knowledge state.

• Recognize the challenges involved in learner modeling and potential solutions.

2.3 The Role of the Learner Model

The learner model is essential for:

• Personalization: Tailoring content and feedback to the learner’s needs.

• Progress Tracking: Monitoring learner development over time.

• Predictive Analytics: Anticipating future performance and identifying potential difficulties.

• Decision Making: Informing the selection of instructional interventions.

2.4 Knowledge Representation in the Learner Model

2.4.1 Knowledge Components

Knowledge in the learner model is often broken down into Knowledge Components (KCs), which are the fundamental units of knowledge or skills that a learner is expected to master.

• Examples of KCs:

  • Mathematical concepts like fractions, algebraic expressions.

  • Language skills like grammar rules, vocabulary words.

  • Procedural skills like laboratory techniques.

2.4.2 Cognitive Structures

• Concept Maps: Graphical representations of concepts and their interrelations.

• Skill Hierarchies: Trees or networks indicating prerequisite relationships among skills.

• Schemas: Structured clusters of knowledge representing complex ideas.

2.5 Mathematical Modeling of Learner Knowledge

2.5.1 Bayesian Knowledge Tracing (BKT)

Overview:

BKT is a probabilistic framework used to model the acquisition of discrete skills over time.

Assumptions:

• Each KC is binary: either mastered or not mastered.

• The learner has a probability of transitioning from unmastered to mastered with each practice opportunity.

• There is a probability of making a slip (incorrect answer despite mastery) or guess (correct answer despite lack of mastery).

Parameters:

• $P(L_0)$: Initial probability that the learner has mastered the KC.

• $P(T)$: Probability of learning the KC after an opportunity.

• $P(S)$: Probability of slipping.

• $P(G)$: Probability of guessing.

Mathematical Formulation:

1. Update of Mastery Probability Before Observing Response:

$P(L_t) = P(L_{t-1}) + [1 - P(L_{t-1})] \cdot P(T)$

2. Update After Observing Response $O_t$:

   • If the response is correct ($O_t = 1$):

$P(L_t \mid O_t) = \frac{P(L_t) \cdot (1 - P(S))}{P(L_t) \cdot > (1 - P(S)) + [1 - P(L_t)] \cdot P(G)}$

• If the response is incorrect ($O_t = 0$):

$P(L_t \mid O_t) = \frac{P(L_t) \cdot P(S)}{P(L_t) \cdot > P(S) + [1 - P(L_t)] \cdot (1 - P(G))}$

Example Calculation:

Suppose:

• $P(L_0) = 0.2$

• $P(T) = 0.1$

• $P(S) = 0.05$

• $P(G) = 0.2$

After a correct response ($O_1 = 1$):

1. Update before observing response:

2. Update after observing response:

So, the probability that the learner has mastered the KC increases to approximately 64.88%.

2.5.2 Extensions of BKT

• Contextual Factors: Incorporating factors like time of day, affective states.

• Individualization: Adapting model parameters per learner.

2.5.3 Item Response Theory (IRT)

IRT models the probability of a correct response as a function of the learner’s ability and item characteristics.

2.5.3.1 The Rasch Model (1PL IRT)

• $\theta$: Learner’s ability.

• $\beta_i$: Difficulty parameter of item $i$.

2.5.3.2 The 2-Parameter Logistic Model (2PL IRT)

• $\alpha_i$: Discrimination parameter of item $i$.

• Higher $\alpha_i$ means the item better discriminates between learners of different abilities.

2.5.3.3 Estimating Ability

• Maximum Likelihood Estimation (MLE):

Find $\hat{\theta}$ that maximizes the likelihood of observed responses.

• Bayesian Estimation:

Incorporate prior distributions over $\theta$ and compute posterior estimates.

2.6 Updating the Learner Model

2.6.1 Sequential Bayesian Updating

As new evidence (responses) is observed, update the learner’s ability estimate $\theta$ using Bayes’ theorem.

• Posterior Probability:

$P(\theta \mid X) \propto P(X \mid \theta) \cdot P(\theta) >$

• Incremental Updating:

After each response, update the posterior distribution of $\theta >$.

2.6.2 Incorporating Time Dynamics

• Forgetting Models:

Account for the decay of knowledge over time.

• Exponential Decay Model:

$P(L_t) = P(L_0) \cdot e^{- \lambda t}$

• $\lambda$: Forgetting rate.

• Spaced Repetition Algorithms:

Schedule reviews based on predicted forgetting.

2.7 Handling Multiple Skills

2.7.1 Multidimensional IRT

Model multiple abilities simultaneously.

• Ability Vector:

$\mathbf{\theta} = [\theta_1, \theta_2, \dots, \theta_K] >$

• Item Response Function:

$P(X_i = 1 \mid \mathbf{\theta}) = \frac{1}{1 + e^{- > (\sum_{k=1}^K \alpha_{ik} \theta_k - \beta_i)}}$

• $\alpha_{ik}$: Discrimination parameter of item $i$ with respect to ability $k$.

2.7.2 Cognitive Diagnosis Models (CDMs)

Provide detailed profiles of learner’s mastery of specific attributes.

• Deterministic Inputs, Noisy “And” Gate (DINA) Model:

  • Items require a subset of attributes.

  • The probability of a correct response depends on mastery of required attributes.

• Mathematical Formulation:

$P(X_i = 1 \mid \mathbf{\alpha_i}, \mathbf{q_i}) = (1 - > s_i)^{\prod_{k=1}^K \alpha_{ik}^{q_{ik}}} g_i^{1 - > \prod_{k=1}^K \alpha_{ik}^{q_{ik}}}$

• $\mathbf{\alpha_i}$: Learner’s attribute mastery vector.

• $\mathbf{q_i}$: Item’s required attribute vector.

• $s_i$: Slip parameter.

• $g_i$: Guessing parameter.

2.8 Challenges in Learner Modeling

2.8.1 Data Sparsity

• Issue: Limited data per learner or skill makes parameter estimation difficult.

• Solutions:

  • Regularization: Penalize model complexity to prevent overfitting.

  • Hierarchical Models: Share statistical strength across similar learners or items.

2.8.2 Non-Stationarity

• Issue: Learner’s ability and engagement can change over time.

• Solutions:

  • Time-Adaptive Models: Incorporate time as a variable.

  • Contextual Bandits: Adapt to changing environments.

2.8.3 Modeling Affective States

• Issue: Emotions like frustration or boredom impact learning.

• Solutions:

  • Affective Computing: Use sensors or interaction patterns to infer affective states.

  • Incorporate into Learner Model: Adjust instructional strategies accordingly.

2.8.4 Ethical Considerations

• Privacy: Ensure compliance with data protection regulations.

• Fairness: Avoid bias against any group of learners.

• Transparency: Make modeling decisions understandable.

2.9 Case Study: Adaptive Language Learning Platform

Background:

A language learning platform aims to personalize vocabulary exercises based on each learner’s proficiency.

Implementation:

• Knowledge Representation:

  • Each word is considered a KC.

  • Words are grouped into themes and difficulty levels.

• Learner Model:

  • Uses BKT to model mastery of each word.

  • Incorporates forgetting by updating mastery probabilities over time.

• Adaptive Algorithm:

  • Selects words for practice based on lowest mastery probabilities.

  • Adjusts for the forgetting curve by scheduling reviews.

Mathematical Application:

• Forgetting Function:

$P(L_t) = P(L_{t-1}) \cdot e^{- \lambda (t - t_{last})}$

• $t_{last}$: Time of last practice.

• $\lambda$: Forgetting rate estimated per learner.

Outcomes:

• Improved Retention: Learners retain vocabulary longer.

• Engagement: Personalized challenges maintain learner interest.

• Feedback: Real-time feedback on mastery helps learners focus their efforts.

2.10 Conclusion

The learner model is a foundational element in adaptive learning systems, enabling the personalization of content and instruction based on an accurate and dynamic representation of each learner’s knowledge and skills. By employing mathematical frameworks like Bayesian Knowledge Tracing and Item Response Theory, adaptive systems can make informed decisions to enhance learning outcomes.

Understanding and effectively implementing learner models require a deep appreciation of both the theoretical underpinnings and practical considerations, including handling uncertainty, data sparsity, and ethical concerns.

2.11 Further Reading

• Corbett, A. T., & Anderson, J. R. (1994). Knowledge tracing: Modeling the acquisition of procedural knowledge. User Modeling and User-Adapted Interaction, 4(4), 253—278.

• van der Linden, W. J. (Ed.). (2016). Handbook of Item Response Theory: Models. CRC Press.

• de la Torre, J. (2011). The generalized DINA model framework. Psychometrika, 76(2), 179—199.

• Pavlik Jr, P. I., Cen, H., & Koedinger, K. R. (2009). Performance factors analysis—a new alternative to knowledge tracing. In Proceedings of the 14th International Conference on Artificial Intelligence in Education (pp. 531—538).

Next Steps

In Lesson 3, we will explore the Domain Model, focusing on how knowledge domains are structured within adaptive learning systems. We will examine ontologies, concept hierarchies, and the mathematical representations that enable effective mapping between the domain content and the learner model.

End of Lesson 2