Abstract
This research paper addresses the significant problem of identifying students-at-risk of failing or dropping out in educational settings. While extensively studied, improving early detection of students likely to fail or drop out remains essential for universities to provide support resources. Previous methods have relied on the students' academic performance to analyse and predict learning strategies, but the complexity of implementing predictive models is heightened by various factors influencing student outcomes. Notably, learning is both a dynamic and socially regulated process, with time and social interactions playing key roles for academic achievement. Nonetheless, despite the importance of these elements, educational research investigating their combined effect is scarce. As grade distribution is affected by several elements, including teaching modalities, grading policies, and course design, identifying students-at-risk and their learning strategies is generally an imbalanced problem, which can lead to biases in predictive algorithms. Our work addresses this issue from a social and temporal perspective, guided by two research questions: (1) To what extent is it possible to inform the early identification of students-at-risk of failing based on interaction data from online discussion forums?, and (2) How does the classification performance compare between traditional oversampling methods and oversampling methods that take the structure of the interactions into account? We based our research on data from an undergraduate course's online forum to build a temporal network of students' communication events across the 12 weeks of the course. Temporal sequences of centrality measures from these interactions served as input for time series classification algorithms. Two oversampling methods are compared: baseline minority oversampling, and a state-of-the-art graph oversampling method that accounts for network structure. Our results show that a temporal network approach, coupled with node oversampling, can enhance student-at-risk identification. However, due to the complexity of the problem and the interactions' sparsity the classification performance is limited when relying solely on this data. We discuss the impact of our findings and contributions, implications, limitations, and future research directions.
