Generative artificial intelligence (GenAI) has the potential as a powerful tool to enhance and strengthen student learning experiences. Students depend on GenAI for immediate answer generating rather than cultivating genuine reflection. To fill this gap, the researchers created a retrieval-augmented generation (RAG)-enabled chatbot, Professor E, which assists students in reflection using Gibbs’s reflective cycle. The researchers used an exploratory methodology to assess students’ AI literacy, reflective capabilities, and their impressions of the system’s educational design. This exploratory study examines students’ AI literacy and reflective capacity, as well as their impressions of the system’s educational design. The researchers gathered data through AI-student conversations to discern the principal themes arising from collaborations between students and AI in an entrepreneurship course in a Hong Kong business school. The focus group interview results highlighted students’ concerns on technical, practical, critical, and ethical factors when utilising Professor E for reflective purposes. The implementation of protected systems, role-playing, and RAG creation helped students in articulating personal thoughts, fostering empathy, and obtaining relevant, customised guidance linked with their project endeavours. This work advances the field of AI-human collaboration by highlighting how pedagogically built GenAI systems can improve students’ reflective thinking abilities.
In recent years, machine learning has become increasingly important for analyzing academic performance in higher education, as it helps identify patterns and supports data-driven decision-making to improve student outcomes. This study presents a machine learning approach to identify and classify academic performance profiles within a higher-education institution. First, unsupervised learning methods are applied to a real-world dataset to detect groups of students with similar characteristics. Next, supervised learning models classify the students into these profiles and are evaluated using traditional classification metrics. In addition, an interpretability analysis is performed based on the importance of characteristics to identify the variables that most strongly influence predictions of performance profiles. We also propose using a Large Language Model to generate human-readable explanations from feature importance outputs. This approach not only evaluates predictive performance but also improves understanding of the key factors that influence outcomes. Experimental results indicate that academic variables have a limited impact on predicting academic performance profiles, whereas institutional variables are more influential, offering valuable insights for decision-making in educational settings.
Traditional think-aloud methods for assessing computational thinking (CT) are labor-intensive and lack real-time feedback. This study developed the “Programming-Assisted Learning App” (PALA), which integrates mobile voice recognition and Sentence-BERT (SBERT) to automate cognitive process analysis. PALA captures students’ verbalized logic, converts it into semantic vectors, and calculates similarity against instructor-defined concepts using cosine similarity. Students receiving scores below a threshold are instantly provided with remedial materials. Preliminary implementation with 60 university students demonstrates that PALA’s similarity scores accurately reflect the mastery of CT elements, such as abstraction and algorithmic thinking. The system successfully reduced text coding time from days to seconds, enabling immediate instructional diagnosis. These findings validate the feasibility of using automated semantic analysis to provide personalized feedback and help instructors dynamically adjust teaching strategies in large-scale programming classrooms.
The growth of digital education has underscored the need to understand and classify the infrastructure that supports eLearning systems. This study classifies eLearning infrastructure by distinguishing between its physical and non-physical components, which together form the foundation of effective online education. A systematic literature review was conducted across major databases (Scopus, Web of Science, IEEE Xplore, and Google Scholar), focusing on studies published between 2020 and 2025. From 300 identified records, 50 duplicates were removed, leaving 250 unique studies for title and abstract screening. Following full-text review and quality appraisal, 30 studies met the inclusion criteria and were included in the final synthesis. Findings were grouped into four key infrastructure categories: (1) Hardware components, (2) Software components, (3) User Experience (UX) Design, and (4) Technical Support and Maintenance. These categories are organised into a two-domain framework: physical infrastructure (hardware) and non-physical infrastructure (software, UX, and support services). A separate discussion addresses components with hybrid characteristics, such as cloud storage. The proposed classification provides a structured basis for evaluating institutional digital readiness and investigating the relationship between infrastructure and the quality of eLearning at tertiary institutions. The findings show that non-physical infrastructure, particularly software components, was the most frequently cited category, appearing in 73
The effectiveness of recommender systems is limited by the cold-start problem, which occurs when insufficient data are available during early use. In educational environments, this issue is amplified by small learner groups, intermittent interaction, and data-collection constraints that delay personalization. This paper proposes a self-adjustable fuzzy inference approach for early-stage personalization in educational recommender systems under cold-start conditions. Unlike traditional fuzzy systems with fixed membership functions, the approach dynamically adapts fuzzy set ranges based on contextual parameters derived from learning activities, enabling interpretable estimation of student model dimensions from the first interactions. The approach was deployed in higher-education STEM courses within an existing recommender system and evaluated across two research cycles in a real-world environment. Results indicate that early-stage adaptive feedback increases acceptance of personalized recommendations and enables a higher proportion of students to reach minimum performance thresholds while preserving stability of summative assessment outcomes.
