Education & Training
Use Case Description
Goal of the use case
Education and training are most effective when the learning tasks (e.g. solving a math problem or assembling a product) are tailored to the skill levels and motivation of individual learners. In order to tailor learning tasks (a.o. the content, complexity, or amount of learning support offered) to the level of the individual learner, we need fine-grained information on learning processes (e.g. how difficult a task is for a certain learner), which is often not readily available or visible to teachers and trainers. Moreover, a current trend in both educational and corporate training contexts is that cohorts of learners (e.g. classes of students, or groups of learners enrolled in a Massive Open Online Course) are becoming more and more heterogeneous, due to both demographic evolutions (e.g. migration) and an increasing need for upskilling and reskilling in the workforce (OECD Skills Outlook, 2021). As a result, teachers and trainers are increasingly faced with a cognitive overload that prevents them from providing quality education, a basic human right and one of the sustainable development goals of the United Nations (SDG4).
Meanwhile, digital technologies are generating a wealth of highly fine-grained data on learning processes in realistic settings (e.g. accuracy of individual responses, how much time and attention learners spend on a task, amount of stress experienced during the task), but the potential of digital data and analytics for providing more effective instruction is not sufficiently exploited. In particular, current techniques cannot deal well with the multidimensional and non-linear nature of learning (for example, recommender systems can only take into account learners’ interest at a particular point in time, not their changing skill levels), and ignore the critical role teachers play in mediating the learning process of their learners.
Therefore, the goal of this use case is to leverage real-world data and develop state-of-the-art artificial intelligence techniques to deliver adaptive education, training and assessment (at scale) in diverse ‘classrooms’ to support effective and efficient learning, while empowering teachers or trainers with actionable information as well as intelligent and explainable decision support systems.
Research within Flanders AI Research Program
Meeting this overall goal requires developing novel AI algorithms and tools that address the following research challenges. We ground each of the challenges in a realistic context by developing a proof-of-concept (PoC) implementation that demonstrates the potential of our AI algorithms and tools.
Research challenge: explainable recommender systems to personalize learning tasks at the level of individual learners, based not just on interest (motivation), but also on skill levels (cognition) and metacognition (PoC 1);
Research challenge: decision support for teachers that visualizes federated data, with a view to adding a human-in-the-loop approach to adaptive education (PoC 2a);
Research challenge: genAI tools for teachers to generate learning materials by exploiting patterns in learners’ previous interactions with the materials (PoC 2b);
Research challenge: automated assessment of learners’ unstructured (spoken) responses (PoC 3).
Industrial and societal relevance
In order to ensure the relevance of our research for the Flemish EdTech ecosystem, we will base our proofs-of-concept on data from real-life applications that emerged from Flemish research projects, such as i-Learn and Nedbox. We also involve educational technology (EdTech) companies (through industry association EdTech Station) as well as end users.
Do you want to learn more or be involved?
Reach out to our use case lead or AI expert lead.
Results, highlights & research outputs
Research results and demonstrators, with publications
PentaForest: an explainable learning activity and trajectory recommender system (PoC1)
PentaForest, a recommender system based on local and global tree-ensemble methods in a multi-target setting, is proposed for recommending learning materials and learning trajectories in the i-Learn dataset. Experimental results demonstrate that the proposed system outperforms the state-of-the-art deep learning-based recommender system (recommender system based on multi-variational Autoencoders) by approximately 10% on two educational datasets w.r.t. ranking performance measures. Due to its tree-based structure, PentaForest offers higher explainability merits compared to deep learning counterparts in modeling learners’ preferences.
Publication
Ilídio, P., Gharahighehi, A., Nakano, F. K., & Vens, C. (2022). Personalized Learning in K-12 Education: Exploring Weak-Labels for a Random Forest-based Collaborative Filtering Approach. In Proceedings of the fist workshop on adaptive lifelong learning (ALL 2024), Recife, Brazil
Enhancing MOOC Recommendations by exploiting preference and engagement personalization signals (PoC1)
Massive Open Online Courses (MOOCs) are emerging as a popular alternative to traditional education, offering learners the flexibility to access a wide range of courses from various disciplines, anytime and anywhere. Despite this accessibility, dropout from MOOCs is significant. To enhance learner engagement, it is crucial to recommend courses that align with their preferences and needs. Time-to-dropout and time-to-completion in MOOCs, like other time-to-event prediction tasks, can be effectively modeled using survival analysis (SA) methods. We applied SA methods to improve collaborative filtering recommendation performance by considering time-to-event in the context of MOOCs. Our proposed approach demonstrates superior performance (~60% boosted performance) compared to collaborative filtering methods trained based on learners' interactions with MOOCs. The findings demonstrate the value of integrating SA methods, trained on engagement data, with recommender systems to enhance personalization in MOOCs.
Publication
Gharahighehi, A., Ghinis, A., Venturini, M., Cornillie, F., & Vens, C. (2025). Enhancing Collaborative Filtering-Based Course Recommendations by Exploiting Time-to-Event Information with Survival Analysis. arXiv preprint arXiv:2503.00072.
Personalizing language learning based on preference, engagement and proficiency (PoC1)
This study investigated AI-driven personalized learning on the free learning platform for Dutch as a second language Nedbox, considering three personalization signals, namely preference, engagement and proficiency. Teachers recommended tailoring homepage content to learners’ preference while filtering out materials with high dropout risk and initializing exercises at a difficulty level matched to the learner’s proficiency. Deep knowledge tracing algorithms outperformed statistical models for proficiency prediction, though they were less explainable. Traditional recommender systems were most effective for content suggestions, and machine learning-based survival analysis methods better modeled engagement compared to traditional approaches. A strong correlation was found between proficiency and engagement, but no significant link between preference and the other two dimensions.
[DEMO] GenAI as a teacher assistant to annotate learning activities (Poc2a)
A tool leveraging generative AI was developed to assist teachers in annotating learning activities with relevant competencies. Its development followed a co-design methodology, involving a series of collaborative activities to gather feedback and iteratively refine the tool in partnership with educators. The system generates predictions regarding the competencies addressed by each activity, based on the nature of its exercises. These predictions, along with accompanying explanations, are produced through a chain of large language models. The tool was enhanced through multiple rounds of teacher feedback, and a qualitative study involving ten educators was conducted to assess the extent to which the generated explanations contribute to the development of user trust.
[DEMO] Supporting teachers in automated question-generation with feedback, control or explanations (Poc2a)
Artificial Intelligence (AI) is becoming more common in classrooms, offering promising ways to improve education. However, challenges remain, especially around giving users control and helping them understand how AI makes decisions. Studies show that when users can interact with and understand AI systems, trust and satisfaction increase, but educational technology has lacked clear guidance in this area. To address this, a tool was developed that automatically generates multiple-choice distractors to help teachers create assessments. It also includes user-centered features: teachers can provide feedback by rating or suggesting distractors, adjust algorithm settings for more control, and view visual explanations to understand how the AI works. These additions aim to make AI tools more effective, transparent, and accessible for educators.
Publication
Szymanski, M., Ooge, J., De Croon, R., Vanden Abeele, V., & Verbert, K. (2024, March). Feedback, control, or explanations? Supporting teachers with steerable distractor-generating AI. In Proceedings of the 14th Learning Analytics and Knowledge Conference (pp. 690-700).
Innovative automatic speech recognition (ASR) modeling for spoken response assessment in the context of language learning (PoC3)
Current end-to-end ASR (Automatic Speech Recognition) systems map speech directly to text. To achieve high performance, these models are trained on huge datasets, ranging from hundreds of thousands to millions of hours of audio, making them highly tolerant to pronunciation variations. This can lead to "wishful hearing," where mispronounced words are still recognized, or grammatical errors are automatically corrected (e.g., "he work" transcribed as "he works"). Similarly, state-of-the-art models like OpenAI’s Whisper tend to omit hesitations and repetitions. In language learning contexts, such behavior is undesirable. Therefore, we aim to reintroduce phonetic and phonemic levels into ASR systems to enhance explainability—allowing the inferred word hypotheses to be interpreted in terms of their actual pronunciation. We focus on training a phoneme recognizer, which will later be integrated into a word recognizer. A key challenge in phoneme recognition is the lack of clear ground truth: training data is labeled orthographically (i.e., with written words), while spoken words can have multiple valid pronunciations. To address this, we extend the popular CTC (Connectionist Temporal Classification) training method to account for pronunciation uncertainty. This approach yields up to a 10% relative reduction in phoneme recognition error rate compared to the reported state-of-the-art error.
Publication
Grafe, H., & Van hamme, H., (2025). Graph Connectionist Temporal Classification for Phoneme Recognition. In Proceedings of IEEE Automatic Speech Recognition and Understanding Workshop. (To appear)
Outputs & Resources
Publications
Ilídio, P., Gharahighehi, A., Nakano, F. K., & Vens, C. (2022). Personalized Learning in K-12 Education: Exploring Weak-Labels for a Random Forest-based Collaborative Filtering Approach. In Proceedings of the fist workshop on adaptive lifelong learning (ALL 2024), Recife, Brazil
Gharahighehi, A., Ghinis, A., Venturini, M., Cornillie, F., & Vens, C. (2025). Enhancing Collaborative Filtering-Based Course Recommendations by Exploiting Time-to-Event Information with Survival Analysis. arXiv preprint arXiv:2503.00072.
Szymanski, M., Ooge, J., De Croon, R., Vanden Abeele, V., & Verbert, K. (2024, March). Feedback, control, or explanations? Supporting teachers with steerable distractor-generating AI. In Proceedings of the 14th Learning Analytics and Knowledge Conference (pp. 690-700).
Grafe, H., & Van hamme, H., (2025). Graph Connectionist Temporal Classification for Phoneme Recognition. In Proceedings of IEEE Automatic Speech Recognition and Understanding Workshop. (To appear)
Gharahighehi, A., Van Schoors, R., Topali, P., & Ooge, J. (2025). The 2nd Workshop on Adaptive Lifelong Learning (ALL25). LAK25 Companion Proceedings, 471-474.
Gharahighehi, A., Van Schoors, R., Topali, P., & Ooge, J. (2024, July). Adaptive lifelong learning (all). In International Conference on Artificial Intelligence in Education (pp. 452-459)
Van Schoors, R., Gharahighehi, A., Fastré, A. (2024). The potential of AI in education and learning. Presented at the Learning Bytes Festival 2024, Kortrijk, Belgium.
Datasets
i-Learn MyWay is a classroom portal for personalized digital learning in K12 education. The dataset contains logs of 10K students from a total of 3K learning activities and 1K learning tracks, anonymized using state-of-the-art privacy-preserving techniques.
Nedbox is a platform to learn Dutch as a second language, coordinated by the Centre for Language and Education at KU Leuven. The dataset contains 57K sessions of 11k students engaged in 600 publications (articles and videos) and 14K exercises.
Demonstrators / Applications
PentaForest: an explainable learning activity and trajectory recommender system.
GenAI as a teacher assistant to annotate learning activities.
Supporting teachers in automated question-generation with feedback, control or explanations.
Innovative automatic speech recognition modeling for spoken response assessment in the context of language learning.
Outreach
The second workshop on Adaptive Lifelong Learning at LAK conference (Gharahighehi, A., Van Schoors, R., Topali, P., & Ooge, J. (2025). The 2nd Workshop on Adaptive Lifelong Learning (ALL25). LAK25 Companion Proceedings, 471-474.)
The first workshop on Adaptive Lifelong Learning at AIED conference (Gharahighehi, A., Van Schoors, R., Topali, P., & Ooge, J. (2024, July). Adaptive lifelong learning (all). In International Conference on Artificial Intelligence in Education (pp. 452-459))
Learning Bytes Festival 2024 (Van Schoors, R., Gharahighehi, A., Fastré, A. (2024). The potential of AI in education and learning. Presented at the Learning Bytes Festival 2024, Kortrijk, Belgium.)
Team & Contact
Contact
Use case lead: Frederik Cornillie (Itec, an imec research group at KU Leuven)
AI expert lead: Alireza Gharahighehi (Itec, an imec research group at KU Leuven)
Use Case Lead(s)
Frederik Cornillie, Itec, an imec research group at KU Leuven
Principal Investigators / AI Expert Leads
Alireza Gharahighehi, Itec, an imec research group at KU Leuven