Massive Open Online Courses (MOOCs) have changed learning by providing high-quality learning materials to the masses in open-source formats. Their very high dropout rates, though, constitute a significant disadvantage, and predicting methods are necessary to identify potentially non-completing students. This paper is an extensive comparative analysis of self-labeled algorithm strategies for the prediction of early certification in MOOCs. Self-labeling algorithms, that leverage semi-supervised learning to refine predictions based on sparsely labeled data, are especially well adapted to the sparsity and skew typical of MOOC data collections. We compare a range of self-labeling approaches, including Self-Training, Co-Training, and Tri-Training, on a variety of MOOC data collections, considering learner participation and interaction patterns. The results give important insights into the strengths and weaknesses of each algorithm, with practical recommendations for their application in MOOC platforms. Our findings indicate that the application of self-labeled methods with particular tailoring can improve early certification prediction, enabling timely interventions and enhancing learner retention.