针对模糊C均值聚类算法需要给定聚类数目、对初值敏感的缺点，本文结合部分监督思想改进了传统的模糊C均值聚类算法。首先对数据进行模糊C均值聚类，然后通过分析划分矩阵中各行的数据特点，选出清晰类，并将其类中心作为标签数据保存用以指导最终的聚类；同时结合划分矩阵中各列的数据特点，评价聚类结果的有效性，调整类数，对剔除清晰类的数据再次进行模糊聚类。重复上述过程至达到循环终止条件，最终根据标签数据进行原始数据的部分监督模糊聚类，得到符合数据内在结构的聚类结果。本文将改进的算法在3个人工数据集和1个真实数据集上进行试验，并与传统模糊C均值聚类结果对比，结果表明改进后的算法有效地避免了初值敏感，得到了较为自然的聚类结果。

Aiming to reduce the sensitivity to the initial cluster centers of fuzzy C-means clustering and avoid determining the number of clusters in advance, a method based on the partially supervised fuzzy clustering is proposed in this paper. First, the data is clustered with fuzzy C-means clustering. Then each row of the partition matrix is analyzed to select the clear clusters, and save their centers as knowledge for the clustering. Combining the analysis of each column and row of the partition matrix, the clustering result is evaluated to adjust the number of clusters. Next, the unclear clusters are clustered again with the adjusted number. This process is repeated until the terminating condition is reached. Finally, use the partially supervised fuzzy clustering to cluster the original data with those saved centers. The method is tested on 3 self-defined datasets and a real dataset, with comparison to the traditional fuzzy C-means clustering results. It is shown that the improved method is insensitive to the initial value and can get an appropriate clustering result.