摘要
距离度量对模糊聚类算法FCM的聚类结果有关键性的影响。实际应用中存在这样一种场景,聚类的数据集中存在着一定量的带标签的成对约束集合的辅助信息。为了充分利用这些辅助信息,首先提出了一种基于混合距离学习方法,它能利用这样的辅助信息来学习出数据集合的距离度量公式。然后,提出了一种基于混合距离学习的鲁棒的模糊C均值聚类算法(HR-FCM算法),它是一种半监督的聚类算法。算法HR-FCM既保留了GIFP-FCM(Generalized FCM algorithm with improved fuzzy partitions)算法的鲁棒性等性能,也因为所采用更为合适的距离度量而具有更好的聚类性能。实验结果证明了所提算法的有效性。
The distance metric plays a vital role in the fuzzy C-means clustering algorithm. In actual applications, there is a practical scenario in which the clustered data have a certain amount of side information, such as pairwise constraints with labels. To sufficiently utilize this side information, first, we propose a learning method based on hybrid distance, in which side information can be utilized to attain a distance metric formula for the data set. Next, we propose a robust fuzzy C-means clustering algorithm （HR-FCM algorithm） based on hybrid-distance learning, which is semi-supervised. The HR-FCM inherits the robustness of the GIFP-FCM （ generalized FCM algorithm with improved fuzzy partitions） and has better clustering performance due to the more appropriate distance metric. The experimental results confirm the effectiveness of the proposed algorithm.
作者
卞则康
王士同
BIAN Zekang WANG Shitong(School of Digital Media, Jiangnan University, Wuxi 214122, China)
出处
《智能系统学报》
CSCD
北大核心
2017年第4期450-458,共9页
CAAI Transactions on Intelligent Systems
基金
国家自然科学基金项目(61272210)
