Abstract: Multi-label feature selection aims to select representative attributes in multi-label scenarios.Most of the existing multi-label feature selection methods focus on obtaining all the feature spaces in advance without considering the streaming feature situation. These features constantly flow into the model one by one over time. In addition, other streaming feature methods need to specify parameters before learning.Therefore, before training different types of data sets, how to select uniform and optimal parameters becomes a difficult problem..Motivated by this, this paper defines the adaptive neighborhood rough set relationship-Gap, and proposes the Multi-Label Online stream Feature Selection based on Adaptive Neighborhood Rough Set(ML-OFS-ANRS).The data mining of neighborhood rough sets does not require any prior knowledge of the feature space structure. It also does not breakingthe neighborhood and order structure of the data when dealing with mixed data.In the first stage, relevant and importantfeatures are selected into the selected subset based on dynamic maximal-dependency. To filter redundant features, the importance of each feature is calculated and parallel reduction is performed in the selected subsetas the second stage.Thus, with the "dynamic maximal-dependency, online irrelevancy discarding"evaluation criteria, ML-OFS-ANRS can select features with high correlation and low redundancy.Experimental results show that ML-OFS-ANRS is superior to traditional feature selection methods and advanced online stream feature selection algorithms when the number of features is the same on 10 different types of data sets.