2026 ASEE Annual Conference & Exposition

Collaborative learning, the process of students engaging with one another to construct knowledge, has consistently been shown to enhance learning outcomes. Instructors encourage this interaction through various in-person and online activities designed to promote discussion and teamwork. However, analyzing collaboration in physical classrooms remains challenging, as in-person interactions are difficult to capture and quantify. Moreover, in-person and online collaborations differ fundamentally in nature, making it difficult to translate findings from digital environments such as learning management systems (LMS) or massive open online courses (MOOCs) to face-to-face classrooms.

Social learning networks (SLNs), which are computational graphs in which nodes represent students and links (i.e., edges) denote interactions between students, have been widely used to model student interactions. Prior work has successfully applied SLNs to online environments to perform link prediction (i.e., predicting when a pair of students are likely to interact in an effort to optimize collaborative learning activities), where interaction data can be easily collected and stored. However, few studies have extended these analyses to in-person classrooms. Furthermore, even across online courses, the ways in which SLNs are formed (e.g., how links are defined, what data are available, and how student interactions are structured) vary substantially. This diversity makes it difficult to use a single analytical framework across multiple courses, where SLNs from a particular classroom are known to better inform link prediction in other courses. Moreover, centralizing such sensitive student data in a location outside the university where the data is collected and stored is prohibited by the Family Educational Rights and Privacy Act (FERPA), which restricts sharing identifiable student information across institutions.

In this work, we jointly address each of these challenges by developing a federated link prediction framework that enables the training of feature based feed forward neural networks (FNNs) to perform link prediction in SLNs from multiple courses, each with different definitions of student interaction, while preserving data privacy. Our approach is designed to be independent of course modality, supporting both in-person and online classrooms. Specifically, we use the Federated Averaging (FedAvg) algorithm, a method where each entity (i.e., a course or institution) locally hosts its own data and trains a FNN on its locally stored dataset. Periodically, the FNN parameters (rather than the data itself) are centralized and aggregated to synchronize the link prediction behaviors learned at each entity. The updated FNN models are returned to their entities where model training resumes. This iterative process enables the model to learn from patterns across diverse courses while maintaining strict student privacy compliance.

We evaluate our framework on four MOOC datasets, two in-person discussion-forum based dataset, and one in-person electrical engineering course reflecting natural variance in how links are represented across courses. Results show that our federated feature-based FNN approach achieves stronger link prediction performance than conventional baselines. The framework effectively learns from data distributed across heterogeneous courses and is especially useful in early stages of a course when SLNs are sparse and few links have formed or in courses with lower engagement where many links do not naturally form. By providing a versatile, privacy-preserving framework, this work supports instructors in identifying potential collaborations, improving group formation, and fostering more effective learning environments across diverse educational settings.