| Massive Open Online Courses (MOOCs) have revolutionized education, offering open access to quality learning resources. However, their high dropout rates pose a significant challenge, necessitating predictive methods to identify learners at risk of non-completion. This paper conducts a comprehensive comparative analysis of self-labeled algorithms in predicting early certification outcomes in MOOCs. Self-labeled algorithms, which leverage semi-supervised learning to enhance predictions from limited labeled data, are well-suited for the sparsity and imbalance often present in MOOC datasets. We evaluate multiple self-labeled techniques, including Self-Training, Co-Training, and Tri-Training, on diverse MOOC datasets, considering factors such as learner engagement and interaction patterns. The results reveal critical insights into the strengths and limitations of each algorithm, providing actionable recommendations for their deployment in MOOC platforms. Our findings demonstrate that tailored use of self-labeled methods can improve early certification prediction, enabling timely interventions and fostering learner retention. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.