Predicting student achievement through peer network analysis for timely personalization via generative AI

Texto completo

Background: Peer influence is a significant determinant in shaping students' academic performance, yet it is often overlooked in traditional educational strategies. The ability to analyze peer influence and collaboration is an important piece in personalizing student educational experiences. Objective: This study aims to investigate how peer interactions can be used to predict students' achievement levels and create clusters based on these predictions. Building on these clusters, a novel AI-driven approach is introduced to personalize the learning experiences for each group, providing content aligned with students' predicted achievement levels and ensuring that students receive the right resources throughout their entire education. Methods: Bi-weekly surveys were conducted throughout the academic year with 45 students enrolled in a computer science program, where students nominated up to five peers they collaborated with the most. The data were used to construct a weighted directed graph (also known as directed network), a type of directed graph allowing multiple edges between the same nodes. Centrality measures were calculated from this directed graph to classify students into low-achieving and high-achieving groups for two-class classification, and into low-achieving, high-achieving, and neutral groups for three-class classification. Generative AI was then utilized to create tailored learning scenarios for each student cluster. For high achievers, the scenario includes providing advanced topics, critical-thinking quizzes and supplementary materials such as research papers and real-world projects. For low achievers, these materials include foundational course materials, multiple-choice quizzes and supplementary materials like videos and step-by-step visualizations. The neutral group was given the option to request specific learning materials and self-assessment quizzes on demand to address their individual learning needs. Results: Eigenvector centrality of the peer collaboration network was identified as the best predictor of student achievement. To assess students' perceived acceptance of the proposed AI-driven personalization, a questionnaire was distributed. The responses indicated that students found the idea appealing and believed it could enhance their learning experience. Several suggestions for improvement were also provided, which will be used to refine the proposed solution in future iterations. Several improvements to the classification mechanism that can enhance its accuracy and effectiveness were also identified during the process. Conclusion: The findings of this research provide educators, administrators, and policymakers with essential information that can be used to build targeted interventions and support systems for students early in their studies. The findings contribute to the broader discourse on educational strategies, enhancing positive peer interactions and developing a learning environment that provides support and motivation to all students.