Abstract: A feature selection method for partial label learning based on neighborhood rough sets is proposed. A partial label neighborhood decision system is constructed, and the concepts of lower approximation and dependency of neighborhood rough sets are then defined in partial label learning. On this basis, a feature selection algorithm suitable to partial label classification is developed. This method can measure the similarity between labels in the set of candidate labels while granulating the feature space in the neighborhood, and select a subset of features with strong relevance to the label information. Two generation mechanisms for false positive candidate labels are used which are different from the most often used random method, and their impact on the results are compared and analyzed in the experiments. Finally, extensive experimental results on six real-world and six controlled synthetic partial label data sets are presented to demonstrate the effectiveness of the proposed feature selection method.

Key words: partial label learning, feature selection, partial label neighborhood decision system, neighborhood rough sets