Rough sets-based tri-trade for partially labeled data

The theory of rough sets is one of the most representative models for handling supervised data entangled with vagueness, impreciseness, or uncertainty. However, little work has been devoted to learning from partially labeled data using rough sets. In this study, a rough sets-based tri-trade model is proposed for partially labeled data. More specifically, a new discernibility matrix that considers both labeled and unlabeled data is first proposed, based on which a beam search-based heuristic algorithm is provided to generate multiple semi-supervised reducts. Then, a tri-trade model using three diverse semi-supervised reducts is developed, in which a data editing technique is embedded to generate reliable pseudo-labels for unlabeled data to improve the tri-trade model. Both theoretical analysis and comparative experiments on the UCI datasets show that the proposed model can effectively utilize unlabeled data to improve generalization performance and compare favorably to other representative methods.