Source-linked AI summary
Exploiting Cognitive Structure for Adaptive Learning
Qi Liu, Shiwei Tong, Chuanren Liu, Hongke Zhao, Enhong Chen, Haiping Ma, Shijin Wang
TL;DR
Adaptive learning needs to model both learners’ evolving knowledge levels and the prerequisite structure of learning items, whereas existing methods often focus on only one. CSEAL combines Knowledge Tracing, Cognitive Navigation, and an Actor-Critic Recommender in an MDP, and experiments report effectiveness and robustness across diverse evaluation settings. Its simulator-based evaluation remains bounded because each simulator has known limitations.
Problem
Existing adaptive-learning methods often model either learners’ evolving knowledge levels or learning-item knowledge structure, although both are important for personalized path recommendation.
Method
CSEAL models sequential recommendation as an MDP, traces knowledge levels with LSTM-based Knowledge Tracing, navigates prerequisite structure to select candidates, and uses an Actor-Critic Recommender.
Results
Experiments with seven baselines, two simulators, diverse scenarios, and human experts demonstrate CSEAL’s effectiveness and robustness.
Takeaways & Limitations
CSEAL jointly exploits learner knowledge levels and item prerequisites to provide dynamically tailored learning paths while maintaining path logicality and reducing search space.
Takeaways & Limitations
The two simulators are imperfect: KSS’s rule-based knowledge evolution may differ from the real world, while KES has difficulty describing knowledge-structure relations.
Abstract
from arXiv · showhide
Adaptive learning, also known as adaptive teaching, relies on learning path recommendation, which sequentially recommends personalized learning items (e.g., lectures, exercises) to satisfy the unique needs of each learner. Although it is well known that modeling the cognitive structure including knowledge level of learners and knowledge structure (e.g., the prerequisite relations) of learning items is important for learning path recommendation, existing methods for adaptive learning often separately focus on either knowledge levels of learners or knowledge structure of learning items. To fully exploit the multifaceted cognitive structure for learning path recommendation, we propose a Cognitive Structure Enhanced framework for Adaptive Learning, named CSEAL. By viewing path recommendation as a Markov Decision Process and applying an actor-critic algorithm, CSEAL can sequentially identify the right learning items to different learners. Specifically, we first utilize a recurrent neural network to trace the evolving knowledge levels of learners at each learning step. Then, we design a navigation algorithm on the knowledge structure to ensure the logicality of learning paths, which reduces the search space in the decision process. Finally, the actor-critic algorithm is used to determine what to learn next and whose parameters are dynamically updated along the learning path. Extensive experiments on real-world data demonstrate the effectiveness and robustness of CSEAL.
1 INTRODUCTION
Adaptive learning personalizes learning paths using learners’ evolving knowledge levels and relationships among learning items. CSEAL combines both cognitive-structure components in sequential recommendation and reports effectiveness and interpretability gains.
- Motivation: Adaptive learning recommends personalized learning items and paths based on each learner’s current knowledge and prerequisite relations.The example path C →D →B →... is tailored to a learner seeking multiplication.
- Challenges: Cognitive structure comprises learners’ evolving, unobservable knowledge levels and the prerequisite-based knowledge structure of learning items.Existing adaptive-learning methods often use only one of these two components.
- Challenges: Knowledge levels affect learning effectiveness, while prerequisite relations constrain paths so learners encounter foundational items before advanced ones.For example, weak mastery of two-digit addition makes multiplication difficult, and prerequisites B and C precede target D.
- CSEAL: CSEAL models path recommendation as an MDP, traces evolving knowledge with an LSTM-based Knowledge Tracing model, and navigates the knowledge structure to select candidates.Its Actor-Critic Recommender updates parameters to improve the whole learning path rather than only one item.
- CSEAL: CSEAL’s experiments show significant improvements over several baselines and provide interpretable insights into learning-path recommendations.The introduction reports effectiveness and interpretability outcomes without giving numerical values here.
2 RELATED WORK
Related work spans generic sequence recommendation, cognitive-structure modeling, and reinforcement learning. Cognitive-structure methods separately model evolving knowledge levels or item relationships, motivating their integration in CSEAL.
- Learning Path Recommendation: Learning-path recommendation has used collaborative filtering and deep sequence models such as KNN, MPR, and GRU4Rec.These methods address learning paths as sequence-recommendation problems.
- Learning Path Recommendation: Cognitive-structure recommendation branches into modeling learners’ knowledge-level evolution and using learning-item knowledge structures.Prior work includes MDP-based knowledge evolution and explicit knowledge-structure-based recommendation.
- Cognitive Structure: Knowledge tracing models learners’ knowledge over time, with DKT using an RNN to represent states continuously in a high-dimensional space.Knowledge graphs separately represent domain entities and relations such as prerequisites.
- Reinforcement Learning: Reinforcement learning refines model parameters from task-specific reward signals and has been applied to dialogue and treatment recommendation.These applications illustrate reinforcement learning for sequential decisions involving complex relationships.
3 PRELIMINARIES
The paper formulates adaptive learning as sessions that recommend sequential learning paths toward target items, using examinations to measure promotion. A prerequisite graph and historical records define the recommendation problem.
- Learning Sessions: A learning session contains a learning path of items and examinations that retrieve the path’s learning effectiveness.Sessions may correspond to homework, chapters, or semesters and have learner- or tutor-defined targets.
- Learning Sessions: Learning effectiveness EP normalizes the change from beginning score Es to ending score Ee by the available improvement to full score Esup.The paper gives EP = (Ee − Es)/(Esup − Es).
- Problem Formulation: A prerequisite graph represents learning items as nodes and prerequisite relations as directed edges, encoding the hierarchical order of learning.Learners generally encounter basic items before more advanced ones.
- Problem Formulation: Given historical records H, target T, and prerequisite graph G, the task is to recommend an N-length path that maximizes whole-path effectiveness while observing each new interaction.Each interaction records the recommended item and the learner’s performance score.
- Problem Formulation: The learning-session formulation treats sequential path generation as a Markov Decision Process with defined states, actions, and rewards.This decision-making formulation supports tailored path policies based on cognitive structure.
4 CSEAL
CSEAL models learning-path recommendation as an MDP whose state combines the learner’s evolving knowledge level with the learning target. It traces knowledge, narrows recommendations through prerequisite-based cognitive navigation, and uses an actor-critic recommender to select subsequent items.
- MDP formulation: CSEAL represents each recommendation step with the learning target and the learner’s current knowledge level, while actions generate recommended learning items.Rewards are received after actions, with the objective of maximizing discounted returns across the learning session.
- Knowledge Tracing: Knowledge Tracing uses an embedded DKT architecture to infer evolving knowledge levels from previous learning records.The records are embedded into lower-dimensional vectors, processed by an LSTM, and converted into knowledge-level representations.
- Cognitive Navigation: Cognitive Navigation selects candidate items from prerequisite-graph neighborhoods that can still reach the learning target, preserving logical path structure.The procedure expands successors and predecessors around a central focus, then removes candidates unable to reach the target.
- Actor-Critic Recommender: The actor samples actions from the candidate set using a policy network, while the critic estimates expected returns from the state.The state combines the traced knowledge level with the target, and actor-critic training updates the policy and value networks.
- Actor-Critic Recommender: CSEAL adds a policy-enhanced loss because overly rapid value-network convergence can slow or prevent policy-network convergence.The resulting loss function includes hyperparameters α and β.
5 EXPERIMENTS
The experiments use learner-session logs from junyiacademy.org together with a prerequisite graph, while recognizing that multiple related concepts may contribute to one session’s learning result.
- Dataset: The dataset contains more than 39 million learner logs, with each record linking a learner, exercise concept, session, correctness, and timestamp.Exercises map one-to-one to knowledge-graph nodes, and records sharing a session id form one learner session.
- Dataset: A prerequisite graph is extracted from the knowledge graph, with edges representing prerequisite relations between concepts.Loops are removed so the resulting prerequisite graph remains a directed acyclic graph.
- Dataset: Session records are grouped by session id and sorted by timestamp to form learning sequences for experiments.These sequences preserve the ordered exercises practiced by the same learner during one session.
- Session structure: Multiple concepts and related concepts may be learned simultaneously during a session, contributing jointly to the final learning result.Figure 5 presents the distributions of sessions in this context.
5.2 System Simulators
The study evaluates learning-path recommendation in two complementary simulators because realistic logs cannot directly provide outcomes for unobserved exercises. KSS models prerequisite effects with rules, while KES learns evolving knowledge levels from data.
- Motivation: Realistic logs contain static exercise sequences and correctness labels, so they cannot directly evaluate or train recommendations for exercises absent from a sequence.This motivates simulated environments for calculating promotion and training agents.
- KSS: KSS is a knowledge-structure-based simulator in which mastery of prerequisite items affects mastery and learning effectiveness on successor items.Its qualitative knowledge-development pattern is designed to fit the prerequisite structure.
- KES: KES trains a DKT model on existing records to estimate each learner’s current knowledge level and uses exercise-answer probabilities as mastery values.KES requires a learning record to initialize the learner’s original knowledge level.
- Simulator limitations: The two simulators are complementary but imperfect: KSS may differ from real-world knowledge evolution, whereas KES may inadequately describe knowledge-structure relations.The framework is evaluated across both environments to test robustness under their differing limitations.
5.3 Experimental Setup
The experimental setup uses the two simulators as environments for online training and evaluation, with simulator-specific data partitions, model dimensions, and regularization settings.
- Environment: Both simulators serve as environments for evaluating and online-training the learning-path recommendation models.Their different characteristics require distinct data partitioning and preprocessing procedures.
- KSS setup: KSS generates an offline dataset of 4,000 records with maximum length 50, split into training, validation, and testing sets at 80/10/10.Its prerequisite graph has 10 nodes and 12 links, with randomly selected learning targets.
- Evaluation: Table 3 reports average EP for all models as the primary evaluation measure.The table summarizes overall performance across the compared approaches.
- KES setup: KES initializes each learner with the first 60% of a session, masks the middle 20%, and reserves the final 20% as the learning target.The resulting datasets are also divided into training, validation, and testing sets at 80/10/10.
- Implementation: The simulators use different model scales: KSS has 10 learning items, while KES has 835, requiring larger embedding, LSTM, output, and value-policy dimensions for KES.Training uses Xavier initialization, minibatches of 16, dropout, and gradient clipping.
5.4 Baseline Approaches
The experiments compare CSEAL with similarity-based, session-based, reinforcement-learning, Monte Carlo, and cognitive-navigation baselines implemented for the learning-path recommendation task.
- Comparison set: The study compares CSEAL against several methods to assess its effectiveness and robustness.The comparison includes both conventional recommendation models and approaches using reinforcement learning or cognitive structure.
- Conventional baselines: KNN recommends the next item using nearby learners identified by cosine distance between their learning paths.It selects a predefined number of nearest learners and uses their behavior for recommendation.
- Conventional baselines: GRU4Rec is a session-based recommender that maps the session sequence to a probability distribution over items likely to appear next.It represents a classical sequence-based baseline.
- Reinforcement-learning baselines: MCS combines Monte Carlo Search with Knowledge Tracing and ranks search paths using KT-predicted promotion.DQN-based methods use reinforcement learning but require substantial domain knowledge to design the MDP transition matrix and exact initial state.
5.5 Evaluation Metrics
Learning path recommendation evaluates learning effectiveness rather than item-selection accuracy, making classical recommendation metrics unsuitable. Quantifying learning effects remains challenging.
- Learning path recommendation focuses on learning effectiveness rather than selecting practiced items or correcting individual exercises.
- Classical metrics such as precision, recall, and NDCG cannot be applied directly to learning path recommendation.
- The quantification of learning effect remains unclear and challenging.
5.6 Experimental Results
Experiments evaluate CSEAL’s promotion, logicality, training stability, path length, and qualitative recommendations across simulated learning settings. CSEAL generally performs best, while cognitive navigation supports logical paths and reduces search difficulty.
- Experimental setup: 20 learning items are used per recommended path, based on the session-length median reported from the data.
- Promotion comparison: CSEAL achieves the best average EP among the evaluated models by combining knowledge-level modeling with Cognitive Navigation.
- Expert comparison: CSEAL outperforms all baselines in expert evaluations of learning-path logicality.
- Expert comparison: Cognitive Navigation helps maintain learning-path logicality according to expert scores.
- Training stability: Cognitive Navigation enables CSEAL to use knowledge structure to obtain an optimal policy stably and reduces the reinforcement-learning search space.
- Path-length analysis: In KSS, expected promotion grows with path length, whereas effectiveness in KES can hardly improve after length 20.
- Path visualization: For a learner struggling with completing_the_square_1, CSEAL recommends reviewing prerequisite items before returning to the target.
6 CONCLUSIONS
The paper concludes that CSEAL combines knowledge tracing, Cognitive Navigation, and actor-critic recommendation for adaptive learning. Experiments across simulators and expert evaluations demonstrate effectiveness and robustness, while the framework’s components remain open to improvement.
- CSEAL retrieves learner knowledge levels, maintains item knowledge structure, and dynamically recommends learning items during a learning cycle.
- Experiments with seven baselines across two simulators, diverse scenarios, and human experts demonstrate CSEAL’s effectiveness and robustness.
- Each step of CSEAL may be further improved in future work.