2026 ASEE Annual Conference & Exposition

Collaborative learning, where students actively engage with their peers to exchange ideas, solve problems, and deepen understanding, has consistently been shown to enhance retention, critical thinking, and overall academic achievement. Recognizing these benefits, many instructors intentionally design activities to increase both the frequency and the quality of peer-to-peer interaction. However, despite the central role of collaboration in student success, methods for systematically modeling and analyzing in-person classroom interactions remain scarce. Social learning networks (SLNs) provide a powerful computational framework for representing collaborative behavior, where students are represented as nodes and their interactions form weighted edges. Prior SLN research, however, has focused almost exclusively on virtual environments such as massive open online courses (MOOCs) or learning management system (LMS) discussion forums, where online posts and clicks create rich, automatically logged data. In these digital settings, researchers have successfully applied network-based analyses, such as link prediction, which is the forecasting of future connections between students based on current network structure, to understand learning dynamics and optimize group formation. Despite the promise of these approaches, no prior work has demonstrated how SLNs or link prediction can be used to capture the complexity of a physical classroom, where face-to-face interactions are not automatically recorded and the absence of quantified data presents a fundamental modeling challenge. To address this challenge, we develop a framework for constructing and analyzing SLNs in a physical classroom that addresses the central challenge of modeling in-person student links. Specifically, we collect weekly student feedback on their experience with collaborative learning exercises and use these evaluations to form weighted computational graphs of peer interactions throughout the term. These graphs capture both the frequency and quality of collaboration, enabling the construction of a dynamic SLN that reflects evolving classroom relationships in the physical classroom. We then implement a graph neural network (GNN) approach for link prediction using a modified GraphSAGE architecture capable of handling weighted edges, representing the satisfaction of students working with their collaborative partners, extending beyond the binary links used in standard implementations. Our framework is evaluated in a real-world electrical engineering course at an American R1 institution. Empirical evaluation demonstrates the viability of our method, achieving a final link-prediction AUC of 0.8000 and outperforming multiple state-of-the-art baselines. By demonstrating that in-person student interactions can be effectively modeled as a weighted SLN and leveraged for accurate link prediction, this work establishes a foundation for data-driven support of collaborative learning in physical classrooms. The proposed framework provides instructors with actionable insights to identify and recommend beneficial student pairings, opening new opportunities to enhance academic outcomes through informed network design.