中心约束的跨源学习可能性C均值聚类算法

Central-constraints possibilistic C-means algorithms based on source domain

可能性C均值聚类算法(Possibilistic C-Means,PCM)相比于模糊C均值聚类算法(Fuzzy C-Means,FCM),能更好地处理含有噪音和例外点的数据,但在处理数据粘性较强的数据集时,PCM算法的聚类中心趋于一致,从而导致聚类算法直接失效。针对这个问题,提出了中心约束准则与跨域迁移学习准则,并将其应用到可能性C均值算法中,从而提出一种具有中心约束能力的聚类算法,简称中心约束的跨源学习聚类算法,改进后的算法能够利用跨域知识进行辅助聚类,确保类中心相互远离,从而能够保证算法的聚类性能。通过模拟数据集和真实数据集的实验,验证了该算法的上述优点。

Compared with Fuzzy C-Means(FCM), Possibilistic C-Means clustering algorithm(PCM)can deal with the data with noise and exception point better, but when dealing with the data set with strong viscosity, the clustering center of PCM algorithm will lead to the direct failure of clustering algorithm. To solve the above issue, this paper devises centralconstraints and transfer based on source domain criterions, and applies these to PCM. It proposes Central-Constraints Possibilistic C-Means algorithms based on the Source Domain(CCSD_PCM for short), which can achieve better clustering effect. Improved algorithm can use the cross-domain knowledge to support the clustering, so as to guarantee the clustering performance of the algorithm. Through the simulation data sets and real data sets, it verifies the above-mentioned advantages of the algorithm.