Understanding the intelligence behind adaptive in adaptive educational systems

The world is undergoing a major transformation which involves rapid advancements in technology like artificial intelligence. Adaptive educational systems (AES) are intelligent systems that personalize learning experiences based on each student’s individual needs, behavior, and progress. These systems go beyond static content delivery. Adaptive intelligence combines artificial intelligence (AI) with human thinking to create learning experiences that are unique to each student. This article explores the foundational intelligence behind AES, including their core components, operational mechanisms, and the interplay between human cognition and machine learning.

Traditional educational models employ a one-size-fits-all approach, that overlooks individual differences in the learners’ abilities, preferences, and learning trajectories. This inflexibility hinders the development of crucial competencies like critical thinking and creativity, leaving graduates unprepared for a rapidly changing work environment (Chigbu et al., 2023). Adaptive educational systems (AES) address this limitation by dynamically adjusting instructional content and strategies based on real-time analysis of the learner. The real-time analyses are based on machine learning, natural language processing, and data analytics that are effective at detecting subtle trends in the learner’s performance data (Chigbu et al., 2024). This also means that the system can monitor topics that a student struggles with and provide help in the form of explanations. This makes learning more active and engaging, rather than just passively receiving information.

Core Functionalities of the Adaptive Technology in AES

Continuous monitoring and analysis of student interactions is a key component of the adaptive educational systems (AES) that facilitates personalized instructions. Such real-time performance data forms the backbone of educational recommender systems and guiding decisions on what question, concept, or module to present. Gligorea et al. (2023), found that a variety of ML algorithms, such as K-means clustering, decision trees, and reinforcement learning, are being used in Adaptive educational systems (AES) to personalize the learning experience. However, the studies that used standard methods like logistic regression and decision trees for predictions, did not consider the effect of time on student learning. This is crucial because students’ knowledge naturally improves over time. Incorporating domain knowledge, specifically student and course features, is highly beneficial for the recommendation system (Elbadrawy & Karypis, 2016). Thai-Nghe et al. (2025), successfully demonstrated that recommender system techniques, specifically matrix factorization and tensor factorization, are highly effective for predicting student performance. Matrix factorization is a dimensionality reduction technique commonly used in collaborative filtering. It decomposes a large student content interaction matrix into two low-rank matrices representing latent features of students and items. The parameters of this model are optimized to minimize the root mean squared error, often using stochastic gradient descent, and a regularization term to prevent overfitting.

Recommender systems are known to help e-commerce users to find interesting items by predicting their preferences. They are widely used in e-commerce and have recently been applied in AES to predict student performance. The recommender systems in AES focusses on generating a high-quality rating score rather than explicitly recommending learning object s. These scores go beyond simple right-or-wrong outcomes to reflect deeper aspects of learning behavior. The rating scores are categorical representations of how useful, appropriate, or effective a learning item is for a particular learner. These scores are used as:

• Labels for supervised learning models
• Ground truth in collaborative filtering
• Targets for matrix/tensor factorization

Traditional ratings rely on binary accuracy (correct/incorrect) or feedback. However, these can be limited or noisy in learning contexts (Thai-Nghe et al., 2011).

Labeling based on such nuanced dynamic assessments leads to more precise learner models, increasing the prediction accuracy of recommended content or interventions (Papamitsiou & Economides, 2014). Dynamic assessments also helps to differentiate between high-performing but low-effort learners and struggling but engaged ones, allowing recommendations to align with learning needs rather than just performance statistics (Heffernan & Heffernan, 2014).

Furthermore, Knowledge Tracing is a family of models that track and estimate a learner’s evolving mastery of skills or concepts over time based on their interactions with the educational content. It operates on sequential learning data, using each new response to update its prediction of what the learner knows and what they’re ready to learn next (Corbett & Anderson, 1994). Common Knowledge Tracing techniques used for dynamic assessmens include:

These models update the probability that a learner has mastered a skill after each practice attempt, incorporating hints used, errors made, and timing. Knowledge Tracing when integrated with dynamic assessment data, enables AES to disaggregate learner profiles more precisely than traditional performance scores.

What Makes a System Truly Adaptive?

In the field of machine learning, “truly adaptive” refers to systems that can continuously adjust, learn, and improve their behavior and predictions in real-time without requiring manual intervention. Unlike traditional machine learning models that depend on periodic retraining, truly adaptive models self-regulate based on streaming data and changing goals. The data from the real-world evolves over time, where the underlying patterns or relationships in the data change over time in unexpected ways which is known as concept drift (Lu et al., 2018). A truly adaptive model detects and adjusts to concept drift automatically to maintain accuracy. Truly adaptive systems leverage incremental learning or online learning approaches that allow models to learn one instance at a time and update themselves with each new observation. These systems use heuristic adaptations of existing batch learning systems to update models as the environment changes (Hoi et al., 2021). Online learning approaches are designed for data that arrives in a sequential order, where the system continuously learns and updates its predictions. Unlike traditional data, which can be stored and processed all at once data streams require algorithms that can learn and make predictions in real-time without storing all past information. While traditional models are good for structured data and clear relationships, modern deep learning models excel at handling complex, unstructured data but require large datasets.

Algorithms such as adaptive random forests or dynamic ensemble models are designed to detect concept drifts and adapt to it (Gomes et al., 2017). The Adaptive Random Forest (ARF) approach adapts to concept drift through several key methods:

• Drift Adaptation Strategy: Each tree in the ARF ensemble has its own “drift monitor.” When a “warning” of a potential change in data patterns is detected for a tree, a new “background tree” starts training in parallel. If the warning escalates to a full “drift” the old tree is replaced by its newly trained background tree. This prevents a sudden drop in performance.
• Theoretically Sound Resampling: ARF utilizes a theoretically sound resampling method based on online bagging. This allows for training trees in parallel.
• Flexible Drift Detection: ARF avoids being tied to a specific drift detection algorithm. It uses both ADWIN and Page Hinkley Test, allowing for future adaptations.
• Background Training and Replacement: When a warning is detected, new trees are trained in the background and then replace existing trees when a drift is detected.
• Weighted Voting: ARF also incorporates weighted voting as part of its strategies to cope with concept drifts.

An ARF-powered recommender system in AES can dynamically adjust recommendations based on the ongoing performance, engagement, and feedback from each student.

Moreover, AES generate high-velocity data like clickstream, quiz scores, time-on-task etc. ARF’s stream-learning design allows continuous learning from this data without needing to batch-train the model. This is essential in truly adaptive systems, where the learner’s journey is non-linear and personalized (Levin, 2024). The benefits of using machine learning in adaptive learning are clear:

• it can make learning truly personal,
• help students with different needs,
• make the learning process more efficient, and
• even predict when a student might struggle so help can be offered early.

Adaptation Techniques in Adaptive Educational Systems (AES)

AES use various adaptation techniques and approaches that dynamically modify the learning environment based on learner profiles, behaviors, and performance. Human learning is influenced by a wide range of variables, including:

• Cognitive abilities like prior knowledge and working memory.
• Affective states like motivation, boredom and anxiety.
• Behavioral traits like perseverance and gaming the system.
• Learning preferences like visual vs verbal.
• Cultural and contextual factors like language and socioeconomic status.

A single adaptation strategy is insufficient to address this multidimensional learner diversity (Brusilovsky & Millán, 2007). Furthermore, learners do not progress in a linear fashion. Their understanding fluctuates based on:

• Fatigue, motivation, or emotional state
• Sudden concept clarity or confusion
• External circumstances

Multiple adaptation types like behavioral, affective, and context-aware are essential to maintain relevance and effectiveness throughout this fluctuating journey (D’mello & Graesser, 2013). Using multiple adaptation strategies enables the system to reduce cognitive overload and enhance long-term retention through personalized sequencing. Studies show that integrated adaptation across content, feedback, and assessment leads to better outcomes than static or single-layer systems (Matsuda et al., 2015; Van Der Linden & Glas, 2010). Major adaptation techniques used in AES are:

• Learner Model-Based Adaptation: This is the foundation of most AES. A learner model stores individual characteristics such as prior knowledge, cognitive styles, learning goals, affective states, and performance history.
• Content Adaptation Techniques: It modifies the learning material presentation, difficulty, or sequencing, to fit the learner’s profile.
• Assessment Adaptation: Also known as adaptive testing, this technique tailors the assessment process by adjusting the type and difficulty of questions.
• Behavioral Adaptation: AES track and respond to learner behavior such as time on task, clicks, hesitations, or help requests.
• Affective Adaptation: Incorporates the learner’s emotional state into adaptation strategies. For instance, systems may detect frustration, boredom, or confidence through sensor data or facial analysis.

No single method can holistically accommodate the dynamic cognitive, emotional, and environmental variables that influence how students learn. These adaptation techniques interact with machine learning algorithms like Adaptive Random Forest (ARF) to power question recommender systems in an Adaptive Educational Systems (AES). The integration demonstrates a strong potential for achieving truly adaptive learning systems. The evolution of Adaptive Educational Systems (AES) represents a significant leap toward achieving personalized, scalable, and effective digital learning experiences. Such systems do not simply recommend content; they orchestrate the learner’s journey by adapting to moment-by-moment interactions. This capability holds immense potential for improving academic outcomes, learner engagement, and self-regulated learning in a scalable and data-driven manner.

References

• Brusilovsky, P., & Millán, E. (2007). User Models for Adaptive Hypermedia and Adaptive Educational Systems. In P. Brusilovsky, A. Kobsa, & W. Nejdl (Eds.), The Adaptive Web: Methods and Strategies of Web Personalization (pp. 3–53). Springer. https://doi.org/10.1007/978-3-540-72079-9_1
• Chigbu, B. I., Ngwevu, V., & Jojo, A. (2023). The effectiveness of innovative pedagogy in the industry 4.0: Educational ecosystem perspective. Social Sciences & Humanities Open , 7 (1), 100419. https://doi.org/10.1016/j.ssaho.2023.100419
• 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. https://doi.org/10.1007/BF01099821
• D’mello, S., & Graesser, A. (2013). AutoTutor and affective autotutor: Learning by talking with cognitively and emotionally intelligent computers that talk back. ACM Trans. Interact. Intell. Syst. , 2 (4), 23:1-23:39. https://doi.org/10.1145/2395123.2395128
• Elbadrawy, A., & Karypis, G. (2016). Domain-Aware Grade Prediction and Top-n Course Recommendation. Proceedings of the 10th ACM Conference on Recommender Systems , 183–190. https://doi.org/10.1145/2959100.2959133
• Gomes, H. M., Bifet, A., Read, J., Barddal, J. P., Enembreck, F., Pfharinger, B., Holmes, G., & Abdessalem, T. (2017). Adaptive random forests for evolving data stream classification. Machine Learning , 106 (9), 1469–1495. https://doi.org/10.1007/s10994-017-5642-8
• Heffernan, N. T., & Heffernan, C. L. (2014). The ASSISTments Ecosystem: Building a Platform that Brings Scientists and Teachers Together for Minimally Invasive Research on Human Learning and Teaching. International Journal of Artificial Intelligence in Education , 24 (4), 470–497. https://doi.org/10.1007/s40593-014-0024-x
• Hoi, S. C. H., Sahoo, D., Lu, J., & Zhao, P. (2021). Online learning: A comprehensive survey. Neurocomputing , 459 , 249–289. https://doi.org/10.1016/j.neucom.2021.04.112
• Levin, S. M. (2024). Review of Machine Learning Models for Application in Adaptive Learning for Higher Education Student. IJFMR – International Journal For Multidisciplinary Research , 6 (2). https://doi.org/10.36948/ijfmr.2024.v06i02.15481
• Lu, J., Liu, A., Dong, F., Gu, F., Gama, J., & Zhang, G. (2018). Learning under Concept Drift: A Review. IEEE Transactions on Knowledge and Data Engineering , 1–1. https://doi.org/10.1109/TKDE.2018.2876857
• Matsuda, N., Cohen, W. W., & Koedinger, K. R. (2015). Teaching the Teacher: Tutoring SimStudent Leads to More Effective Cognitive Tutor Authoring. International Journal of Artificial Intelligence in Education , 25 (1), 1–34. https://doi.org/10.1007/s40593-014-0020-1
• Papamitsiou, Z., & Economides, A. (2014). Learning Analytics and Educational Data Mining in Practice: A Systematic Literature Review of Empirical Evidence. Educational Technology & Society , 17 , 49–64.
• Shute, V. J. (2008). Focus on Formative Feedback. Review of Educational Research , 78 (1), 153–189. https://doi.org/10.3102/0034654307313795
• Thai-Nghe, N., Drumond, L., Horváth, T., Nanopoulos, A., & Schmidt-Thieme, L. (2011). MATRIX AND TENSOR FACTORIZATION FOR PREDICTING STUDENT PERFORMANCE . 69–78. https://www.scitepress.org/Link.aspx?doi=10.5220/0003328700690078
• Van Der Linden, W. J., & Glas, C. A. W. (Eds.). (2010). Elements of Adaptive Testing . Springer. https://doi.org/10.1007/978-0-387-85461-8