2024 ASEE Annual Conference & Exposition

The early prediction of student performance has long been a significant area of interest within the educational research community. Numerous studies have sought to identify students who struggle early in their courses, with the goal of providing timely interventions. This early detection of at-risk students is vital for fostering and promoting student success, which is critical to the missions of higher-education institutions and their long-term goals of improving student graduation and retention rates.

This paper builds on the predictive models from a previous paper, focused on making early course-based predictions of student performance based on data collected from a Learning Management System (LMS). It describes these models, augments them with new features, and investigates the portability and robustness of their predictions— aspects not addressed in the previous paper. It frames the early prediction of student performance as a binary classification problem, distinguishing between two student groups: at-risk students who are failing their courses and require early identification for timely assistance, and those who are performing satisfactorily. Additionally, this paper employs Explainable AI (XAI) methods to gain insight into how the new models generate their predictions and to identify the key data features that contribute to these predictions.

This paper relies on data gathered from the Canvas Learning Management System through RESTful and GraphQL APIs. The data pertain to 274 lower-division Computer Science courses, encompassing 2,656 students and 37 instructors. These courses are delivered in various formats (face-to-face, online, and virtual), spanning a three-year period at a public four-year university.