Abstract
The semi-supervised learning (SSL) is a recent machine learning paradigm where both labeled and unlabeled data are taken into consideration in order to enrich the learning process. In the current work we deal with the robustness of network-based SSL algorithms. For this purpose, statistical analysis with validation criterions are carried out on designed data and simulations. We have observed that the algorithms robustness vary according to the data smoothness, the clustering consistency with the data labels, the labeled data representative ness, and the suitability of the network-construction model to the SSL algorithm. In the lack of these properties, even the best SSL algorithms have their performance depreciated. On the other hand, to consider pair wise constraints for network-construction models based on the labeled data, improves significantly the SSL robustness.