In this study, a novel ROugh set based Semi-supervised fEature Selector (Roses) was developed to pre-process partially labeled data. The main innovations of our Roses are: 1) the selected features over labeled samples laid the foundation for further searching qualified features over unlabeled samples; 2) a new granularity related measure was designed to quickly evaluate features. Through testing four different ratios (20%, 40%, 60%, 80%) of labeled samples, the experimental results over 15 UCI datasets demonstrated that our framework is superior to the other five popular partially labeled data feature selectors: 1) the feature subsets identified by Roses offer competitive classification performances; 2) Roses is good at seeking a balance between efficiency of searching features and effectiveness of the selected features.