<p>The development of multi-label feature selection aims to alleviate the curse of dimensionality that is intrinsic to multi-label learning contexts. Nevertheless, the majority of existing feature selection methods adopt a uniform treatment of all labels, failing to fully account for the similarity existing between labels. However, in practical scenarios, inherent similarities exist between labels—highly similar labels tend to form label groups, which is primarily characterized by their co-occurrence across a large number of samples. This phenomenon leads to a critical limitation: when evaluating feature significance, conventional methods are more inclined to select features strongly correlated with large-scale label groups, while neglecting those that exhibit high relevance to small-scale label groups. Consequently, the performance of such methods is compromised to a certain extent. Additionally, existing methods rarely comprehensively integrate feature–label correlation, label–label correlation, and feature–feature redundancy. Moreover, the adoption of fixed fuzzy neighborhood radius further hinders the performance. To tackle these aforementioned issues, this paper designs a feature selection method that combines the label grouping strategy and the fuzzy multi-neighborhood information measures. Specifically, highly similar labels are first clustered into distinct label groups, and different fuzzy neighborhood radii are assigned to individual features. Subsequently, the comprehensive performance of features on all label groups evaluated through the fuzzy multi-neighborhood information measures. Finally, experimental results obtained from 15 datasets demonstrate that the proposed method delivers better performance than 9 representative multi-label feature selection methods.</p>