Semi-supervised Classification (SSC) is becoming an attractive research filed due to the emergence of real-world problems on which the number of unlabeled examples exceeds the labeled ones. The natural complexity of this kind of problems rices up when designing algorithms with some interpretability features. In order to overcome this challenge, a novel SSC model called Self-labeling Grey-box (SlGb) has been recently proposed. The SlGb algorithm uses a black-box classifier to enlarge the dataset with the unlabeled examples and a white-box to build an interpretable model. In this paper, we attempt boosting the prediction rates of the SlGb algorithm by editing the training set using the knowledge acquired with rough sets. This can be achieved by weighting the instances according to their inclusion degree to rough information granules before building the final, white-box classification model.