在现实世界中，多标记数据特征维度高且样本数量大，因此，从中提取有价值的特征可以有效减少多标记学习算法的计算负担，从而提高运行效率和分类性能。模糊邻域粗糙集是一种特征选择工具，能够从给定的信息系统中获取决策规则。互补信息度量可用于量化信息系统的不确定性。因此，本文针对多标记数据集，将经典互补信息度量推广到模糊邻域粗糙集中研究特征选择问题，主要研究内容如下：

      (1)首先，考虑邻域半径对特征选择结果的影响，定义了模糊邻域互补信息度量，并讨论了其相关性质；其次，引入模糊决策，证明了带有模糊决策的多标记模糊邻域互补条件熵和互补互信息随着特征增加分别具有单调不增和单调不减的性质；最后，依据模糊邻域互补熵和互补互信息内外部特征重要度评价准则的等价性，提出了广义多标记模糊邻域互补熵特征选择算法。
     (2)首先，利用互补熵的概念，构建了度量标记之间辨别关系的差别矩阵，并结合最小生成树原理，将标记分为不同的标记组；其次，将模糊邻域互补信息度量从二元扩展至三元，给出了基于模糊邻域互补互信息的4个相关定义，并对其性质进行了讨论；最后，利用模糊邻域互补互信息、互补条件互信息、互补交互信息量化标记集的特征相关性、交互性和冗余性，提出了一种基于特征交互性的多标记特征选择算法。
      (3)通过对模糊邻域半径和算法终止阈值两个参数的讨论，选择了能够为以上算法提供最佳性能的参数组合。在6个多标记数据集上使用多标记分类器(MLKNN)进行了仿真实验，并与其他相关成熟算法进行对比。实验结果表明，本文提出的两个算法在5个多标记评价指标上的表现明显优于其他算法。同时，对本文的两个算法进行比较，基于特征交互性的多标记特征选择算法性能优于广义多标记模糊邻域互补熵特征选择算法。

     In the real world, the featuredimension of multi-label data is high and the number of samples is large. Therefore, extracting valuable features from multi-label data can effectively reduce the computational burden of multi-label learning algorithms, thereby improving operational efficiency and classification performance. Fuzzy neighborhood rough setis a feature selection tool that can obtain decision rules from a given information system. Complementary information measurecan be used to quantify the uncertainty ofinformation systems. Therefore,this paper extends the classical complementary information measure to the fuzzy neighborhood rough set to study the featureselection problem for muli-label data sets. The main research contents are as follows.

     (1)Firstly, considering the influence of neighborhood radius on feature selection results,the fuzzy neighborhood complementary information measure is defined and its related properties are discussed. Secondly, fuzzy decision is introduced, and it is proved that the multi-label fuzzy neighborhood complementary conditional entropy and complementary mutual information with fuzzy decision have the properties of monotone non-increasing and monotone non-decreasing with the increase offeatures, respectively. Finally, according to the equivalence offuzzy neighborhood complementary entropy and complementary mutual information internal and external feature importance evaluation criteria, a generalized multi-label fuzzy neighbor-hood complementary entropy feature selectionalgorithm is proposed.

      (2)Firstly, by using the concept of complementary entropy, the difference matrix is constructed to measure the discrimination relationship between labels. Combined with the principle ofminimum spanning tree, the labels are divided into different label groups. Secondly,the fuzzy neighborhood complementary information measureis extended from binary to ternary,and four related definitions based on fuzzy neighborhood complementary mutual information are given, and their properties are discussed. Finally, a multi-label feature selection algorithm based on feature interactivity is proposed by using fuzzy neighborhood complementary mutual information, complementary conditional mutual information and complementary interactive information to quantify the feature correlation, interactivity and redundancy of the label set. 

      (3)Through the discussion of the two parameters of the fuzzy neighborhood radius and the algorithm termination threshold, the parameter combination that can provide the best perfor-mance for the above algorithm is selected. The multi-label classifier (MLKNN) is used to conduct simulation experiments on six multi-label datasets, and compared with other related mature algorithms. The experimental results show that the performance of the two algorithms proposed in this paper is significantly better than other algorithms on five multi-label evaluation indicators. At the same time, the two algorithms in this paper are compared. The performance of multi-label feature selection algorithm based on feature interaction is better than that of generalized multi-label fuzzy neighborhood complementary entropy feature selection algorithm.